MyArxiv
Computation and Language 77
☆ VisionThink: Smart and Efficient Vision Language Model via Reinforcement Learning
Recent advancements in vision-language models (VLMs) have improved performance by increasing the number of visual tokens, which are often significantly longer than text tokens. However, we observe that most real-world scenarios do not require such an extensive number of visual tokens. While the performance drops significantly in a small subset of OCR-related tasks, models still perform accurately in most other general VQA tasks with only 1/4 resolution. Therefore, we propose to dynamically process distinct samples with different resolutions, and present a new paradigm for visual token compression, namely, VisionThink. It starts with a downsampled image and smartly decides whether it is sufficient for problem solving. Otherwise, the model could output a special token to request the higher-resolution image. Compared to existing Efficient VLM methods that compress tokens using fixed pruning ratios or thresholds, VisionThink autonomously decides whether to compress tokens case by case. As a result, it demonstrates strong fine-grained visual understanding capability on OCR-related tasks, and meanwhile saves substantial visual tokens on simpler tasks. We adopt reinforcement learning and propose the LLM-as-Judge strategy to successfully apply RL to general VQA tasks. Moreover, we carefully design a reward function and penalty mechanism to achieve a stable and reasonable image resize call ratio. Extensive experiments demonstrate the superiority, efficiency, and effectiveness of our method. Our code is available at https://github.com/dvlab-research/VisionThink.
comment: Code and models are available at https://github.com/dvlab-research/VisionThink
☆ Comparing Apples to Oranges: A Dataset & Analysis of LLM Humour Understanding from Traditional Puns to Topical Jokes
Humour, as a complex language form, is derived from myriad aspects of life, whilst existing work on computational humour has focussed almost exclusively on short pun-based jokes. In this work, we investigate whether the ability of Large Language Models (LLMs) to explain humour depends on the particular humour form. We compare models on simple puns and more complex topical humour that requires knowledge of real-world entities and events. In doing so, we curate a dataset of 600 jokes split across 4 joke types and manually write high-quality explanations. These jokes include heterographic and homographic puns, contemporary internet humour, and topical jokes, where understanding relies on reasoning beyond "common sense", rooted instead in world knowledge regarding news events and pop culture. Using this dataset, we compare the zero-shot abilities of a range of LLMs to accurately and comprehensively explain jokes of different types, identifying key research gaps in the task of humour explanation. We find that none of the tested models (inc. reasoning models) are capable of reliably generating adequate explanations of all joke types, further highlighting the narrow focus of most works in computational humour on overly simple joke forms.
☆ A Survey of Context Engineering for Large Language Models
The performance of Large Language Models (LLMs) is fundamentally determined by the contextual information provided during inference. This survey introduces Context Engineering, a formal discipline that transcends simple prompt design to encompass the systematic optimization of information payloads for LLMs. We present a comprehensive taxonomy decomposing Context Engineering into its foundational components and the sophisticated implementations that integrate them into intelligent systems. We first examine the foundational components: context retrieval and generation, context processing and context management. We then explore how these components are architecturally integrated to create sophisticated system implementations: retrieval-augmented generation (RAG), memory systems and tool-integrated reasoning, and multi-agent systems. Through this systematic analysis of over 1300 research papers, our survey not only establishes a technical roadmap for the field but also reveals a critical research gap: a fundamental asymmetry exists between model capabilities. While current models, augmented by advanced context engineering, demonstrate remarkable proficiency in understanding complex contexts, they exhibit pronounced limitations in generating equally sophisticated, long-form outputs. Addressing this gap is a defining priority for future research. Ultimately, this survey provides a unified framework for both researchers and engineers advancing context-aware AI.
comment: ongoing work; 165 pages, 1401 citations
☆ The Imitation Game: Turing Machine Imitator is Length Generalizable Reasoner
Length generalization, the ability to solve problems of longer sequences than those observed during training, poses a core challenge of Transformer-based large language models (LLM). Although existing studies have predominantly focused on data-driven approaches for arithmetic operations and symbolic manipulation tasks, these approaches tend to be task-specific with limited overall performance. To pursue a more general solution, this paper focuses on a broader case of reasoning problems that are computable, i.e., problems that algorithms can solve, thus can be solved by the Turing Machine. From this perspective, this paper proposes Turing MAchine Imitation Learning (TAIL) to improve the length generalization ability of LLMs. TAIL synthesizes chain-of-thoughts (CoT) data that imitate the execution process of a Turing Machine by computer programs, which linearly expands the reasoning steps into atomic states to alleviate shortcut learning and explicit memory fetch mechanism to reduce the difficulties of dynamic and long-range data access in elementary operations. To validate the reliability and universality of TAIL, we construct a challenging synthetic dataset covering 8 classes of algorithms and 18 tasks. Without bells and whistles, TAIL significantly improves the length generalization ability as well as the performance of Qwen2.5-7B on various tasks using only synthetic data, surpassing previous methods and DeepSeek-R1. The experimental results reveal that the key concepts in the Turing Machine, instead of the thinking styles, are indispensable for TAIL for length generalization, through which the model exhibits read-and-write behaviors consistent with the properties of the Turing Machine in their attention layers. This work provides a promising direction for future research in the learning of LLM reasoning from synthetic data.
☆ Vision-and-Language Training Helps Deploy Taxonomic Knowledge but Does Not Fundamentally Alter It
Does vision-and-language (VL) training change the linguistic representations of language models in meaningful ways? Most results in the literature have shown inconsistent or marginal differences, both behaviorally and representationally. In this work, we start from the hypothesis that the domain in which VL training could have a significant effect is lexical-conceptual knowledge, in particular its taxonomic organization. Through comparing minimal pairs of text-only LMs and their VL-trained counterparts, we first show that the VL models often outperform their text-only counterparts on a text-only question-answering task that requires taxonomic understanding of concepts mentioned in the questions. Using an array of targeted behavioral and representational analyses, we show that the LMs and VLMs do not differ significantly in terms of their taxonomic knowledge itself, but they differ in how they represent questions that contain concepts in a taxonomic relation vs. a non-taxonomic relation. This implies that the taxonomic knowledge itself does not change substantially through additional VL training, but VL training does improve the deployment of this knowledge in the context of a specific task, even when the presentation of the task is purely linguistic.
☆ Social and Political Framing in Search Engine Results
Search engines play a crucial role in shaping public discourse by influencing how information is accessed and framed. While prior research has extensively examined various dimensions of search bias -- such as content prioritization, indexical bias, political polarization, and sources of bias -- an important question remains underexplored: how do search engines and ideologically-motivated user queries contribute to bias in search results. This study analyzes the outputs of major search engines using a dataset of political and social topics. The findings reveal that search engines not only prioritize content in ways that reflect underlying biases but also that ideologically-driven user queries exacerbate these biases, resulting in the amplification of specific narratives. Moreover, significant differences were observed across search engines in terms of the sources they prioritize. These results suggest that search engines may play a pivotal role in shaping public perceptions by reinforcing ideological divides, thereby contributing to the broader issue of information polarization.
comment: Accepted to ICWSM 2026
☆ HapticCap: A Multimodal Dataset and Task for Understanding User Experience of Vibration Haptic Signals
Haptic signals, from smartphone vibrations to virtual reality touch feedback, can effectively convey information and enhance realism, but designing signals that resonate meaningfully with users is challenging. To facilitate this, we introduce a multimodal dataset and task, of matching user descriptions to vibration haptic signals, and highlight two primary challenges: (1) lack of large haptic vibration datasets annotated with textual descriptions as collecting haptic descriptions is time-consuming, and (2) limited capability of existing tasks and models to describe vibration signals in text. To advance this area, we create HapticCap, the first fully human-annotated haptic-captioned dataset, containing 92,070 haptic-text pairs for user descriptions of sensory, emotional, and associative attributes of vibrations. Based on HapticCap, we propose the haptic-caption retrieval task and present the results of this task from a supervised contrastive learning framework that brings together text representations within specific categories and vibrations. Overall, the combination of language model T5 and audio model AST yields the best performance in the haptic-caption retrieval task, especially when separately trained for each description category.
☆ The Generative Energy Arena (GEA): Incorporating Energy Awareness in Large Language Model (LLM) Human Evaluations
The evaluation of large language models is a complex task, in which several approaches have been proposed. The most common is the use of automated benchmarks in which LLMs have to answer multiple-choice questions of different topics. However, this method has certain limitations, being the most concerning, the poor correlation with the humans. An alternative approach, is to have humans evaluate the LLMs. This poses scalability issues as there is a large and growing number of models to evaluate making it impractical (and costly) to run traditional studies based on recruiting a number of evaluators and having them rank the responses of the models. An alternative approach is the use of public arenas, such as the popular LM arena, on which any user can freely evaluate models on any question and rank the responses of two models. The results are then elaborated into a model ranking. An increasingly important aspect of LLMs is their energy consumption and, therefore, evaluating how energy awareness influences the decisions of humans in selecting a model is of interest. In this paper, we present GEA, the Generative Energy Arena, an arena that incorporates information on the energy consumption of the model in the evaluation process. Preliminary results obtained with GEA are also presented, showing that for most questions, when users are aware of the energy consumption, they favor smaller and more energy efficient models. This suggests that for most user interactions, the extra cost and energy incurred by the more complex and top-performing models do not provide an increase in the perceived quality of the responses that justifies their use.
☆ AbGen: Evaluating Large Language Models in Ablation Study Design and Evaluation for Scientific Research ACL 2025
We introduce AbGen, the first benchmark designed to evaluate the capabilities of LLMs in designing ablation studies for scientific research. AbGen consists of 1,500 expert-annotated examples derived from 807 NLP papers. In this benchmark, LLMs are tasked with generating detailed ablation study designs for a specified module or process based on the given research context. Our evaluation of leading LLMs, such as DeepSeek-R1-0528 and o4-mini, highlights a significant performance gap between these models and human experts in terms of the importance, faithfulness, and soundness of the ablation study designs. Moreover, we demonstrate that current automated evaluation methods are not reliable for our task, as they show a significant discrepancy when compared to human assessment. To better investigate this, we develop AbGen-Eval, a meta-evaluation benchmark designed to assess the reliability of commonly used automated evaluation systems in measuring LLM performance on our task. We investigate various LLM-as-Judge systems on AbGen-Eval, providing insights for future research on developing more effective and reliable LLM-based evaluation systems for complex scientific tasks.
comment: ACL 2025
☆ Multi-Agent Synergy-Driven Iterative Visual Narrative Synthesis ACL
Automated generation of high-quality media presentations is challenging, requiring robust content extraction, narrative planning, visual design, and overall quality optimization. Existing methods often produce presentations with logical inconsistencies and suboptimal layouts, thereby struggling to meet professional standards. To address these challenges, we introduce RCPS (Reflective Coherent Presentation Synthesis), a novel framework integrating three key components: (1) Deep Structured Narrative Planning; (2) Adaptive Layout Generation; (3) an Iterative Optimization Loop. Additionally, we propose PREVAL, a preference-based evaluation framework employing rationale-enhanced multi-dimensional models to assess presentation quality across Content, Coherence, and Design. Experimental results demonstrate that RCPS significantly outperforms baseline methods across all quality dimensions, producing presentations that closely approximate human expert standards. PREVAL shows strong correlation with human judgments, validating it as a reliable automated tool for assessing presentation quality.
comment: 22 pages, 7 figures, 3 tables. Submitted to an ACL-style conference
Overview of the TalentCLEF 2025: Skill and Job Title Intelligence for Human Capital Management
Advances in natural language processing and large language models are driving a major transformation in Human Capital Management, with a growing interest in building smart systems based on language technologies for talent acquisition, upskilling strategies, and workforce planning. However, the adoption and progress of these technologies critically depend on the development of reliable and fair models, properly evaluated on public data and open benchmarks, which have so far been unavailable in this domain. To address this gap, we present TalentCLEF 2025, the first evaluation campaign focused on skill and job title intelligence. The lab consists of two tasks: Task A - Multilingual Job Title Matching, covering English, Spanish, German, and Chinese; and Task B - Job Title-Based Skill Prediction, in English. Both corpora were built from real job applications, carefully anonymized, and manually annotated to reflect the complexity and diversity of real-world labor market data, including linguistic variability and gender-marked expressions. The evaluations included monolingual and cross-lingual scenarios and covered the evaluation of gender bias. TalentCLEF attracted 76 registered teams with more than 280 submissions. Most systems relied on information retrieval techniques built with multilingual encoder-based models fine-tuned with contrastive learning, and several of them incorporated large language models for data augmentation or re-ranking. The results show that the training strategies have a larger effect than the size of the model alone. TalentCLEF provides the first public benchmark in this field and encourages the development of robust, fair, and transferable language technologies for the labor market.
☆ QuestA: Expanding Reasoning Capacity in LLMs via Question Augmentation
Reinforcement learning (RL) has become a key component in training large language reasoning models (LLMs). However, recent studies questions its effectiveness in improving multi-step reasoning-particularly on hard problems. To address this challenge, we propose a simple yet effective strategy via Question Augmentation: introduce partial solutions during training to reduce problem difficulty and provide more informative learning signals. Our method, QuestA, when applied during RL training on math reasoning tasks, not only improves pass@1 but also pass@k-particularly on problems where standard RL struggles to make progress. This enables continual improvement over strong open-source models such as DeepScaleR and OpenMath Nemotron, further enhancing their reasoning capabilities. We achieve new state-of-the-art results on math benchmarks using 1.5B-parameter models: 67.1% (+5.3%) on AIME24, 59.5% (+10.0%) on AIME25, and 35.5% (+4.0%) on HMMT25. Further, we provide theoretical explanations that QuestA improves sample efficiency, offering a practical and generalizable pathway for expanding reasoning capability through RL.
comment: 19 pages, 8 figures
☆ Automating Steering for Safe Multimodal Large Language Models
Recent progress in Multimodal Large Language Models (MLLMs) has unlocked powerful cross-modal reasoning abilities, but also raised new safety concerns, particularly when faced with adversarial multimodal inputs. To improve the safety of MLLMs during inference, we introduce a modular and adaptive inference-time intervention technology, AutoSteer, without requiring any fine-tuning of the underlying model. AutoSteer incorporates three core components: (1) a novel Safety Awareness Score (SAS) that automatically identifies the most safety-relevant distinctions among the model's internal layers; (2) an adaptive safety prober trained to estimate the likelihood of toxic outputs from intermediate representations; and (3) a lightweight Refusal Head that selectively intervenes to modulate generation when safety risks are detected. Experiments on LLaVA-OV and Chameleon across diverse safety-critical benchmarks demonstrate that AutoSteer significantly reduces the Attack Success Rate (ASR) for textual, visual, and cross-modal threats, while maintaining general abilities. These findings position AutoSteer as a practical, interpretable, and effective framework for safer deployment of multimodal AI systems.
comment: Working in progress. 22 pages (8+ for main); 25 figures; 1 table
☆ HATS: Hindi Analogy Test Set for Evaluating Reasoning in Large Language Models
Analogies test a model's ability to infer implicit relationships between concepts, making them a key benchmark for evaluating reasoning capabilities. While large language models (LLMs) are widely evaluated for reasoning in English, their abilities in Indic languages remain understudied, limiting our understanding of whether these models generalize across languages. To address this gap, we introduce a new Hindi Analogy Test Set (HATS), comprising 405 multiple-choice questions sourced from Indian government exams. We benchmark state-of-the-art multilingual LLMs using various prompting strategies and introduce a grounded Chain of Thought approach that leverages cognitive theories of analogical reasoning. This approach improves model performance on Hindi analogy questions. Our experiments show that models perform best with English prompts, irrespective of the prompting strategy. Our test set addresses the lack of a critical resource to evaluate LLM reasoning capabilities in Hindi.
☆ Enhancing Cross-task Transfer of Large Language Models via Activation Steering
Large language models (LLMs) have shown impressive abilities in leveraging pretrained knowledge through prompting, but they often struggle with unseen tasks, particularly in data-scarce scenarios. While cross-task in-context learning offers a direct solution for transferring knowledge across tasks, it still faces critical challenges in terms of robustness, scalability, and efficiency. In this paper, we investigate whether cross-task transfer can be achieved via latent space steering without parameter updates or input expansion. Through an analysis of activation patterns in the latent space of LLMs, we observe that the enhanced activations induced by in-context examples have consistent patterns across different tasks. Inspired by these findings, we propose CAST, a novel Cross-task Activation Steering Transfer framework that enables effective transfer by manipulating the model's internal activation states. Our approach first selects influential and diverse samples from high-resource tasks, then utilizes their contrastive representation-enhanced activations to adapt LLMs to low-resource tasks. Extensive experiments across both cross-domain and cross-lingual transfer settings show that our method outperforms competitive baselines and demonstrates superior scalability and lower computational costs.
☆ Automatically assessing oral narratives of Afrikaans and isiXhosa children
Developing narrative and comprehension skills in early childhood is critical for later literacy. However, teachers in large preschool classrooms struggle to accurately identify students who require intervention. We present a system for automatically assessing oral narratives of preschool children in Afrikaans and isiXhosa. The system uses automatic speech recognition followed by a machine learning scoring model to predict narrative and comprehension scores. For scoring predicted transcripts, we compare a linear model to a large language model (LLM). The LLM-based system outperforms the linear model in most cases, but the linear system is competitive despite its simplicity. The LLM-based system is comparable to a human expert in flagging children who require intervention. We lay the foundation for automatic oral assessments in classrooms, giving teachers extra capacity to focus on personalised support for children's learning.
comment: Accepted to SLaTE 2025
☆ GEMMAS: Graph-based Evaluation Metrics for Multi Agent Systems EMNLP
Multi-agent systems built on language models have shown strong performance on collaborative reasoning tasks. However, existing evaluations focus only on the correctness of the final output, overlooking how inefficient communication and poor coordination contribute to redundant reasoning and higher computational costs. We introduce GEMMAS, a graph-based evaluation framework that analyzes the internal collaboration process by modeling agent interactions as a directed acyclic graph. To capture collaboration quality, we propose two process-level metrics: Information Diversity Score (IDS) to measure semantic variation in inter-agent messages, and Unnecessary Path Ratio (UPR) to quantify redundant reasoning paths. We evaluate GEMMAS across five benchmarks and highlight results on GSM8K, where systems with only a 2.1% difference in accuracy differ by 12.8% in IDS and 80% in UPR, revealing substantial variation in internal collaboration. These findings demonstrate that outcome-only metrics are insufficient for evaluating multi-agent performance and highlight the importance of process-level diagnostics in designing more interpretable and resource-efficient collaborative AI systems.
comment: 4 figures, 1 algorithm, 2 tables, 6 pages, under review at EMNLP Industry track 2025
☆ Feature-based analysis of oral narratives from Afrikaans and isiXhosa children
Oral narrative skills are strong predictors of later literacy development. This study examines the features of oral narratives from children who were identified by experts as requiring intervention. Using simple machine learning methods, we analyse recorded stories from four- and five-year-old Afrikaans- and isiXhosa-speaking children. Consistent with prior research, we identify lexical diversity (unique words) and length-based features (mean utterance length) as indicators of typical development, but features like articulation rate prove less informative. Despite cross-linguistic variation in part-of-speech patterns, the use of specific verbs and auxiliaries associated with goal-directed storytelling is correlated with a reduced likelihood of requiring intervention. Our analysis of two linguistically distinct languages reveals both language-specific and shared predictors of narrative proficiency, with implications for early assessment in multilingual contexts.
comment: SLaTE 2025 in Nijmegen, Netherlands
☆ Inverse Reinforcement Learning Meets Large Language Model Post-Training: Basics, Advances, and Opportunities
In the era of Large Language Models (LLMs), alignment has emerged as a fundamental yet challenging problem in the pursuit of more reliable, controllable, and capable machine intelligence. The recent success of reasoning models and conversational AI systems has underscored the critical role of reinforcement learning (RL) in enhancing these systems, driving increased research interest at the intersection of RL and LLM alignment. This paper provides a comprehensive review of recent advances in LLM alignment through the lens of inverse reinforcement learning (IRL), emphasizing the distinctions between RL techniques employed in LLM alignment and those in conventional RL tasks. In particular, we highlight the necessity of constructing neural reward models from human data and discuss the formal and practical implications of this paradigm shift. We begin by introducing fundamental concepts in RL to provide a foundation for readers unfamiliar with the field. We then examine recent advances in this research agenda, discussing key challenges and opportunities in conducting IRL for LLM alignment. Beyond methodological considerations, we explore practical aspects, including datasets, benchmarks, evaluation metrics, infrastructure, and computationally efficient training and inference techniques. Finally, we draw insights from the literature on sparse-reward RL to identify open questions and potential research directions. By synthesizing findings from diverse studies, we aim to provide a structured and critical overview of the field, highlight unresolved challenges, and outline promising future directions for improving LLM alignment through RL and IRL techniques.
☆ From Roots to Rewards: Dynamic Tree Reasoning with RL
Modern language models address complex questions through chain-of-thought (CoT) reasoning (Wei et al., 2023) and retrieval augmentation (Lewis et al., 2021), yet struggle with error propagation and knowledge integration. Tree-structured reasoning methods, particularly the Probabilistic Tree-of-Thought (ProbTree)(Cao et al., 2023) framework, mitigate these issues by decomposing questions into hierarchical structures and selecting answers through confidence-weighted aggregation of parametric and retrieved knowledge (Yao et al., 2023). However, ProbTree's static implementation introduces two key limitations: (1) the reasoning tree is fixed during the initial construction phase, preventing dynamic adaptation to intermediate results, and (2) each node requires exhaustive evaluation of all possible solution strategies, creating computational inefficiency. We present a dynamic reinforcement learning (Sutton and Barto, 2018) framework that transforms tree-based reasoning into an adaptive process. Our approach incrementally constructs the reasoning tree based on real-time confidence estimates, while learning optimal policies for action selection (decomposition, retrieval, or aggregation). This maintains ProbTree's probabilistic rigor while improving both solution quality and computational efficiency through selective expansion and focused resource allocation. The work establishes a new paradigm for treestructured reasoning that balances the reliability of probabilistic frameworks with the flexibility required for real-world question answering systems.
☆ Assessing the Reliability of LLMs Annotations in the Context of Demographic Bias and Model Explanation
Understanding the sources of variability in annotations is crucial for developing fair NLP systems, especially for tasks like sexism detection where demographic bias is a concern. This study investigates the extent to which annotator demographic features influence labeling decisions compared to text content. Using a Generalized Linear Mixed Model, we quantify this inf luence, finding that while statistically present, demographic factors account for a minor fraction ( 8%) of the observed variance, with tweet content being the dominant factor. We then assess the reliability of Generative AI (GenAI) models as annotators, specifically evaluating if guiding them with demographic personas improves alignment with human judgments. Our results indicate that simplistic persona prompting often fails to enhance, and sometimes degrades, performance compared to baseline models. Furthermore, explainable AI (XAI) techniques reveal that model predictions rely heavily on content-specific tokens related to sexism, rather than correlates of demographic characteristics. We argue that focusing on content-driven explanations and robust annotation protocols offers a more reliable path towards fairness than potentially persona simulation.
☆ A Computational Framework to Identify Self-Aspects in Text ACL
This Ph.D. proposal introduces a plan to develop a computational framework to identify Self-aspects in text. The Self is a multifaceted construct and it is reflected in language. While it is described across disciplines like cognitive science and phenomenology, it remains underexplored in natural language processing (NLP). Many of the aspects of the Self align with psychological and other well-researched phenomena (e.g., those related to mental health), highlighting the need for systematic NLP-based analysis. In line with this, we plan to introduce an ontology of Self-aspects and a gold-standard annotated dataset. Using this foundation, we will develop and evaluate conventional discriminative models, generative large language models, and embedding-based retrieval approaches against four main criteria: interpretability, ground-truth adherence, accuracy, and computational efficiency. Top-performing models will be applied in case studies in mental health and empirical phenomenology.
comment: Accepted to ACL SRW 2025
☆ SemCSE: Semantic Contrastive Sentence Embeddings Using LLM-Generated Summaries For Scientific Abstracts
We introduce SemCSE, an unsupervised method for learning semantic embeddings of scientific texts. Building on recent advances in contrastive learning for text embeddings, our approach leverages LLM-generated summaries of scientific abstracts to train a model that positions semantically related summaries closer together in the embedding space. This resulting objective ensures that the model captures the true semantic content of a text, in contrast to traditional citation-based approaches that do not necessarily reflect semantic similarity. To validate this, we propose a novel benchmark designed to assess a model's ability to understand and encode the semantic content of scientific texts, demonstrating that our method enforces a stronger semantic separation within the embedding space. Additionally, we evaluate SemCSE on the comprehensive SciRepEval benchmark for scientific text embeddings, where it achieves state-of-the-art performance among models of its size, thus highlighting the benefits of a semantically focused training approach.
☆ Formalizing Attack Scenario Description: A Proposed Model
Organizations face an ever-changing threat landscape. They must continuously dedicate significant efforts to protect their assets, making their adoption of increased cybersecurity automation inevitable. However, process automation requires formalization of input data. Through this paper, we address this need for processes that use attack scenarios as input. Among these processes, one can mention both the generation of scripts for attack simulation and training purposes, as well as the analysis of attacks. Therefore, the paper's main research contribution is a novel formal model that encompasses the attack's context description and its scenario. It is abstracted using UML class model. Once the description of our model done, we will show how it could serve an upstream attack analysis process. We will show also its use for an automatic generation of attack scripts in the context of cybersecurity training. These two uses cases constitute the second contribution of this present research work.
☆ Rethinking the Embodied Gap in Vision-and-Language Navigation: A Holistic Study of Physical and Visual Disparities ICCV 2025
Recent Vision-and-Language Navigation (VLN) advancements are promising, but their idealized assumptions about robot movement and control fail to reflect physically embodied deployment challenges. To bridge this gap, we introduce VLN-PE, a physically realistic VLN platform supporting humanoid, quadruped, and wheeled robots. For the first time, we systematically evaluate several ego-centric VLN methods in physical robotic settings across different technical pipelines, including classification models for single-step discrete action prediction, a diffusion model for dense waypoint prediction, and a train-free, map-based large language model (LLM) integrated with path planning. Our results reveal significant performance degradation due to limited robot observation space, environmental lighting variations, and physical challenges like collisions and falls. This also exposes locomotion constraints for legged robots in complex environments. VLN-PE is highly extensible, allowing seamless integration of new scenes beyond MP3D, thereby enabling more comprehensive VLN evaluation. Despite the weak generalization of current models in physical deployment, VLN-PE provides a new pathway for improving cross-embodiment's overall adaptability. We hope our findings and tools inspire the community to rethink VLN limitations and advance robust, practical VLN models. The code is available at https://crystalsixone.github.io/vln_pe.github.io/.
comment: Accepted by ICCV 2025
☆ Teach Old SAEs New Domain Tricks with Boosting
Sparse Autoencoders have emerged as powerful tools for interpreting the internal representations of Large Language Models, yet they often fail to capture domain-specific features not prevalent in their training corpora. This paper introduces a residual learning approach that addresses this feature blindness without requiring complete retraining. We propose training a secondary SAE specifically to model the reconstruction error of a pretrained SAE on domain-specific texts, effectively capturing features missed by the primary model. By summing the outputs of both models during inference, we demonstrate significant improvements in both LLM cross-entropy and explained variance metrics across multiple specialized domains. Our experiments show that this method efficiently incorporates new domain knowledge into existing SAEs while maintaining their performance on general tasks. This approach enables researchers to selectively enhance SAE interpretability for specific domains of interest, opening new possibilities for targeted mechanistic interpretability of LLMs.
☆ MRT at IberLEF-2025 PRESTA Task: Maximizing Recovery from Tables with Multiple Steps
This paper presents our approach for the IberLEF 2025 Task PRESTA: Preguntas y Respuestas sobre Tablas en Espa\~nol (Questions and Answers about Tables in Spanish). Our solution obtains answers to the questions by implementing Python code generation with LLMs that is used to filter and process the table. This solution evolves from the MRT implementation for the Semeval 2025 related task. The process consists of multiple steps: analyzing and understanding the content of the table, selecting the useful columns, generating instructions in natural language, translating these instructions to code, running it, and handling potential errors or exceptions. These steps use open-source LLMs and fine-grained optimized prompts for each step. With this approach, we achieved an accuracy score of 85\% in the task.
comment: Accepted as an official challenge paper in the PRESTA: Questions and Answers over Tabular Data shared task at IberLEF 2025, colocated with the 41st SEPLN Conference in Zaragoza, Spain
☆ UniSLU: Unified Spoken Language Understanding from Heterogeneous Cross-Task Datasets
Spoken Language Understanding (SLU) plays a crucial role in speech-centric multimedia applications, enabling machines to comprehend spoken language in scenarios such as meetings, interviews, and customer service interactions. SLU encompasses multiple tasks, including Automatic Speech Recognition (ASR), spoken Named Entity Recognition (NER), and spoken Sentiment Analysis (SA). However, existing methods often rely on separate model architectures for individual tasks such as spoken NER and SA, which increases system complexity, limits cross-task interaction, and fails to fully exploit heterogeneous datasets available across tasks. To address these limitations, we propose UniSLU, a unified framework that jointly models multiple SLU tasks within a single architecture. Specifically, we propose a unified representation for diverse SLU tasks, enabling full utilization of heterogeneous datasets across multiple tasks. Built upon this representation, we propose a unified generative method that jointly models ASR, spoken NER, and SA tasks, enhancing task interactions and enabling seamless integration with large language models to harness their powerful generative capabilities. Extensive experiments on public SLU datasets demonstrate the effectiveness of our approach, achieving superior SLU performance compared to several benchmark methods, making it well-suited for real-world speech-based multimedia scenarios. We will release all code and models at github to facilitate future research.
comment: 13 pages, 3 figures
☆ Probabilistic Soundness Guarantees in LLM Reasoning Chains
In reasoning chains generated by large language models (LLMs), initial errors often propagate and undermine the reliability of the final conclusion. Current LLM-based error detection methods often fail to detect propagated errors because they do not properly account for how earlier errors might corrupt judgments of downstream reasoning. To better detect such propagated errors, we introduce Autoregressive Reasoning Entailment Stability (ARES), a novel probabilistic framework that prevents error propagation by judging each claim based only on previously-assessed sound premises. This inductive method yields a nuanced score for each step and provides certified statistical guarantees of its soundness, rather than a brittle binary label. ARES achieves state-of-the-art performance across four benchmarks (72.1% Macro-F1, +8.2 points) and demonstrates superior robustness on very long synthetic reasoning chains, where it excels at detecting propagated errors (90.3% F1, +27.6 points).
☆ Making Language Model a Hierarchical Classifier and Generator
Decoder-only language models, such as GPT and LLaMA, generally decode on the last layer. Motivated by human's hierarchical thinking capability, we propose that a hierarchical decoder architecture could be built with different layers decoding texts simultaneously. Due to limited time and computationally resources, we choose to adapt a pretrained language model into this form of hierarchical decoder. Language heads of the last layer are copied to different selected intermediate layers, and fine-tuned with different task inputs. By thorough experiments, we validate that these selective intermediate layers could be adapted to speak meaningful and reasonable contents, and this paradigm of hierarchical decoder can obtain state-of-the-art performances on multiple tasks such as hierarchical text classification, classification-guided generation, and hierarchical text generation. This study suggests the possibility of a generalized hierarchical reasoner, pretraining from scratch.
☆ Are Knowledge and Reference in Multilingual Language Models Cross-Lingually Consistent?
Cross-lingual consistency should be considered to assess cross-lingual transferability, maintain the factuality of the model knowledge across languages, and preserve the parity of language model performance. We are thus interested in analyzing, evaluating, and interpreting cross-lingual consistency for factual knowledge. We examine code-mixed coreferential statements conveyed identical knowledge across languages to study cross-lingual knowledge consistency. We use some interpretability approaches to analyze the behavior of a model in cross-lingual contexts, discovering that multilingual models show different levels of consistency, subject to language families, linguistic factors, and a bottleneck in cross-lingual consistency on a particular layer. In addition, we evaluate common strategies aimed at improving multilingual performance to observe whether these strategies can improve knowledge consistency at the same time. While knowledge is not cross-lingual consistency in many cases, code-switching training and cross-lingual word alignment objectives show the most promising results, emphasizing the noteworthiness of cross-lingual alignment supervision and code-switching training for both multilingual performance and cross-lingual consistency enhancement.
☆ Emotional Support with LLM-based Empathetic Dialogue Generation
Emotional Support Conversation (ESC) aims to provide empathetic and effective emotional assistance through dialogue, addressing the growing demand for mental health support. This paper presents our solution for the NLPCC 2025 Task 8 ESC evaluation, where we leverage large-scale language models enhanced by prompt engineering and finetuning techniques. We explore both parameter-efficient Low-Rank Adaptation and full-parameter fine-tuning strategies to improve the model's ability to generate supportive and contextually appropriate responses. Our best model ranked second in the competition, highlighting the potential of combining LLMs with effective adaptation methods for ESC tasks. Future work will focus on further enhancing emotional understanding and response personalization to build more practical and reliable emotional support systems.
☆ Large Language Models' Internal Perception of Symbolic Music
Large language models (LLMs) excel at modeling relationships between strings in natural language and have shown promise in extending to other symbolic domains like coding or mathematics. However, the extent to which they implicitly model symbolic music remains underexplored. This paper investigates how LLMs represent musical concepts by generating symbolic music data from textual prompts describing combinations of genres and styles, and evaluating their utility through recognition and generation tasks. We produce a dataset of LLM-generated MIDI files without relying on explicit musical training. We then train neural networks entirely on this LLM-generated MIDI dataset and perform genre and style classification as well as melody completion, benchmarking their performance against established models. Our results demonstrate that LLMs can infer rudimentary musical structures and temporal relationships from text, highlighting both their potential to implicitly encode musical patterns and their limitations due to a lack of explicit musical context, shedding light on their generative capabilities for symbolic music.
☆ MCPEval: Automatic MCP-based Deep Evaluation for AI Agent Models
The rapid rise of Large Language Models (LLMs)-based intelligent agents underscores the need for robust, scalable evaluation frameworks. Existing methods rely on static benchmarks and labor-intensive data collection, limiting practical assessment. We introduce \oursystemname, an open-source Model Context Protocol (MCP)-based framework that automates end-to-end task generation and deep evaluation of LLM agents across diverse domains. MCPEval standardizes metrics, seamlessly integrates with native agent tools, and eliminates manual effort in building evaluation pipelines. Empirical results across five real-world domains show its effectiveness in revealing nuanced, domain-specific performance. We publicly release MCPEval https://github.com/SalesforceAIResearch/MCPEval to promote reproducible and standardized LLM agent evaluation.
comment: https://github.com/SalesforceAIResearch/MCPEval
☆ PMKLC: Parallel Multi-Knowledge Learning-based Lossless Compression for Large-Scale Genomics Database KDD-25
Learning-based lossless compressors play a crucial role in large-scale genomic database backup, storage, transmission, and management. However, their 1) inadequate compression ratio, 2) low compression \& decompression throughput, and 3) poor compression robustness limit their widespread adoption and application in both industry and academia. To solve those challenges, we propose a novel \underline{P}arallel \underline{M}ulti-\underline{K}nowledge \underline{L}earning-based \underline{C}ompressor (PMKLC) with four crucial designs: 1) We propose an automated multi-knowledge learning-based compression framework as compressors' backbone to enhance compression ratio and robustness; 2) we design a GPU-accelerated ($s$,$k$)-mer encoder to optimize compression throughput and computing resource usage; 3) we introduce data block partitioning and Step-wise Model Passing (SMP) mechanisms for parallel acceleration; 4) We design two compression modes PMKLC-S and PMKLC-M to meet the complex application scenarios, where the former runs on a resource-constrained single GPU and the latter is multi-GPU accelerated. We benchmark PMKLC-S/M and 14 baselines (7 traditional and 7 leaning-based) on 15 real-world datasets with different species and data sizes. Compared to baselines on the testing datasets, PMKLC-S/M achieve the average compression ratio improvement up to 73.609\% and 73.480\%, the average throughput improvement up to 3.036$\times$ and 10.710$\times$, respectively. Besides, PMKLC-S/M also achieve the best robustness and competitive memory cost, indicating its greater stability against datasets with different probability distribution perturbations, and its strong ability to run on memory-constrained devices.
comment: Accepted via KDD-25
☆ Learning Robust Negation Text Representations
Despite rapid adoption of autoregressive large language models, smaller text encoders still play an important role in text understanding tasks that require rich contextualized representations. Negation is an important semantic function that is still not properly captured by such methods, affecting many downstream applications relying on text embeddings. We propose a strategy to improve negation robustness of text encoders, by distilling data from large language models using diverse patterns of negation and hedging. We adopt a standard contrastive learning strategy to finetune a strong BERT-based model, and observe large improvement in negation understanding capabilities while maintaining competitive performance on general benchmarks. In addition, we also show that our method can be adapted to LLMs, leading to improved performance on negation benchmarks.
☆ A Comprehensive Survey of Electronic Health Record Modeling: From Deep Learning Approaches to Large Language Models
Artificial intelligence (AI) has demonstrated significant potential in transforming healthcare through the analysis and modeling of electronic health records (EHRs). However, the inherent heterogeneity, temporal irregularity, and domain-specific nature of EHR data present unique challenges that differ fundamentally from those in vision and natural language tasks. This survey offers a comprehensive overview of recent advancements at the intersection of deep learning, large language models (LLMs), and EHR modeling. We introduce a unified taxonomy that spans five key design dimensions: data-centric approaches, neural architecture design, learning-focused strategies, multimodal learning, and LLM-based modeling systems. Within each dimension, we review representative methods addressing data quality enhancement, structural and temporal representation, self-supervised learning, and integration with clinical knowledge. We further highlight emerging trends such as foundation models, LLM-driven clinical agents, and EHR-to-text translation for downstream reasoning. Finally, we discuss open challenges in benchmarking, explainability, clinical alignment, and generalization across diverse clinical settings. This survey aims to provide a structured roadmap for advancing AI-driven EHR modeling and clinical decision support. For a comprehensive list of EHR-related methods, kindly refer to https://survey-on-tabular-data.github.io/.
☆ Synergy: End-to-end Concept Model
In this paper, we present Synergy, a language model that bridges different levels of abstraction in an end-to-end fashion through a learned routing mechanism. Focusing on low-level linguistic abstraction, we trained our model as a byte-level language model. Our model spontaneously learns to tokenize bytes, producing fewer concept tokens than Byte-level Byte Pair Encoder (BBPE) tokenizers while keeping comparable performance. By comparing with Llama3, we observed an advantage of Synergy under the same model scale and training dataset size. Further studies show that the middle part (the higher abstraction part) of our model performs better when positional encodings are removed, suggesting the emergence of position-independent concepts. These findings demonstrate the feasibility of tokenizer-free architectures, paving the way for more robust and flexible pipelines.
☆ Logit Arithmetic Elicits Long Reasoning Capabilities Without Training
Large reasoning models (LRMs) can do complex reasoning via long chain-of-thought (CoT) involving cognitive strategies such as backtracking and self-correction. Recent studies suggest that some models inherently possess these long reasoning abilities, which may be unlocked via extra training. Our work first investigates whether we can elicit such behavior without any training. To this end, we propose a decoding-time approach, ThinkLogit, which utilizes logits arithmetic (Liu et al., 2024) to tune a target large LM for long reasoning using a substantially smaller model as guider. We then show that we can further boost performance by training the guider model with preference optimization over correct/incorrect reasoning pairs sampled from both the target and guider model -- a setup we refer to as ThinkLogit-DPO. Our experiments demonstrate that ThinkLogit and ThinkLogit-DPO achieve a relative improvement in pass@1 by 26% and 29%, respectively, over four mathematical datasets using the Qwen2.5-32B when guided by R1-Distill-Qwen-1.5B -- a model 21x smaller. Lastly, we show that ThinkLogit can transfer long reasoning skills acquired through reinforcement learning, improving pass@1 by 13% relative compared to the Qwen2.5-32B base model. Our work presents a computationally-efficient method to elicit long reasoning in large models with minimal or no additional training.
☆ Strategy Adaptation in Large Language Model Werewolf Agents
This study proposes a method to improve the performance of Werewolf agents by switching between predefined strategies based on the attitudes of other players and the context of conversations. While prior works of Werewolf agents using prompt engineering have employed methods where effective strategies are implicitly defined, they cannot adapt to changing situations. In this research, we propose a method that explicitly selects an appropriate strategy based on the game context and the estimated roles of other players. We compare the strategy adaptation Werewolf agents with baseline agents using implicit or fixed strategies and verify the effectiveness of our proposed method.
comment: 7 pages, 2 figures
☆ TransEvalnia: Reasoning-based Evaluation and Ranking of Translations
We present TransEvalnia, a prompting-based translation evaluation and ranking system that uses reasoning in performing its evaluations and ranking. This system presents fine-grained evaluations based on a subset of the Multidimensional Quality Metrics (https://themqm.org/), returns an assessment of which translation it deems the best, and provides numerical scores for the various dimensions and for the overall translation. We show that TransEvalnia performs as well as or better than the state-of-the-art MT-Ranker (Moosa et al. 2024) on our own English-Japanese data as well as several language pairs from various WMT shared tasks. Using Anthropic's Claude-3.5-Sonnet and Qwen-2.5-72B-Instruct as the evaluation LLMs, we show that the evaluations returned are deemed highly acceptable to human raters, and that the scores assigned to the translations by Sonnet, as well as other LLMs, correlate well with scores assigned by the human raters. We also note the sensitivity of our system -- as well as MT-Ranker -- to the order in which the translations are presented, and we propose methods to address this position bias. All data, including the system's evaluation and reasoning, human assessments, as well as code is released.
☆ FLEXITOKENS: Flexible Tokenization for Evolving Language Models
Language models (LMs) are challenging to adapt to new data distributions by simple finetuning. This is due to the rigidity of their subword tokenizers, which typically remain unchanged during adaptation. This inflexibility often leads to inefficient tokenization, causing overfragmentation of out-of-distribution domains, unseen languages, or scripts. In this work, we develop byte-level LMs with learnable tokenizers to make tokenization adaptive. Our models include a submodule that learns to predict boundaries between the input byte sequence, encoding it into variable-length segments. Existing tokenizer-free methods train this boundary predictor using an auxiliary loss that enforces a fixed compression rate across the training corpus, introducing a new kind of rigidity. We propose FLEXITOKENS, a simplified training objective that enables significantly greater flexibility during adaptation. Evaluating across multiple multilingual benchmarks, morphologically diverse tasks, and domains, we demonstrate that FLEXITOKENS consistently reduces token over-fragmentation and achieves up to 10\% improvements on downstream task performance compared to subword and other gradient-based tokenizers. Code and data for our experiments will be released at https://github.com/owos/flexitokens
☆ AudioJudge: Understanding What Works in Large Audio Model Based Speech Evaluation
Current speech evaluation suffers from two critical limitations: the need and difficulty of designing specialized systems targeting individual audio characteristics, and poor correlation between automatic evaluation methods and human preferences. This work presents a systematic study of Large Audio Model (LAM) as a Judge, AudioJudge, investigating whether it can provide a unified evaluation framework that addresses both challenges. We systematically explore AudioJudge across audio characteristic detection tasks, including pronunciation, speaking rate, speaker identification and speech quality, and system-level human preference simulation for automated benchmarking. We investigate different prompt engineering strategies, finding that audio concatenation combined with in-context learning significantly improves performance across both audio characteristic detection and human preference simulation tasks. We further introduce a multi-aspect ensemble AudioJudge to enable general-purpose multi-aspect audio evaluation. This method decomposes speech assessment into specialized judges for lexical content, speech quality, and paralinguistic features, achieving up to 0.91 Spearman correlation with human preferences on our system ranking benchmark. Robustness analysis reveals that while LAMs maintain strong performance under acoustic noise, they exhibit significant verbosity and positional biases that require careful mitigation.
☆ AdaptiSent: Context-Aware Adaptive Attention for Multimodal Aspect-Based Sentiment Analysis
We introduce AdaptiSent, a new framework for Multimodal Aspect-Based Sentiment Analysis (MABSA) that uses adaptive cross-modal attention mechanisms to improve sentiment classification and aspect term extraction from both text and images. Our model integrates dynamic modality weighting and context-adaptive attention, enhancing the extraction of sentiment and aspect-related information by focusing on how textual cues and visual context interact. We tested our approach against several baselines, including traditional text-based models and other multimodal methods. Results from standard Twitter datasets show that AdaptiSent surpasses existing models in precision, recall, and F1 score, and is particularly effective in identifying nuanced inter-modal relationships that are crucial for accurate sentiment and aspect term extraction. This effectiveness comes from the model's ability to adjust its focus dynamically based on the context's relevance, improving the depth and accuracy of sentiment analysis across various multimodal data sets. AdaptiSent sets a new standard for MABSA, significantly outperforming current methods, especially in understanding complex multimodal information.
comment: 12 pages (including references), 2 figures (Fig. 1 overview, Fig. 2 hyperparameter sensitivity with two subplots), 6 tables (performance, ablation, dataset stats, case studies, etc.), accepted at ASONAM 2025 (Social Network Analysis and Mining)
♻ ☆ DeFine: Decision-Making with Analogical Reasoning over Factor Profiles ACL 2025
LLMs are ideal for decision-making thanks to their ability to reason over long contexts. However, challenges arise when processing speech transcripts that describe complex scenarios, as they are verbose and include repetition, hedging, and vagueness. E.g., during a company's earnings call, an executive might project a positive revenue outlook to reassure investors, despite uncertainty regarding future earnings. It is crucial for LLMs to incorporate this uncertainty systematically when making decisions. In this paper, we introduce \textsc{DeFine}, a modular framework that constructs probabilistic factor profiles from complex scenarios. It then integrates these profiles with analogical reasoning, leveraging insights from similar past experiences to guide LLMs in making critical decisions in new situations. Our framework separates the tasks of quantifying uncertainty and incorporating it into LLM decision-making. This approach is particularly useful in areas such as consulting and financial deliberation, where making decisions under uncertainty is vital.
comment: Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), Vienna, Austria
♻ ☆ Identifying Task Groupings for Multi-Task Learning Using Pointwise V-Usable Information
The success of multi-task learning can depend heavily on which tasks are grouped together. Naively grouping all tasks or a random set of tasks can result in negative transfer, with the multi-task models performing worse than single-task models. Though many efforts have been made to identify task groupings and to measure the relatedness among different tasks, it remains a challenging research topic to define a metric to identify the best task grouping out of a pool of many potential task combinations. We propose a metric of task relatedness based on task difficulty measured by pointwise V-usable information (PVI). PVI is a recently proposed metric to estimate how much usable information a dataset contains given a model. We hypothesize that tasks with not statistically different PVI estimates are similar enough to benefit from the joint learning process. We conduct comprehensive experiments to evaluate the feasibility of this metric for task grouping on 15 NLP datasets in the general, biomedical, and clinical domains. We compare the results of the joint learners against single learners, existing baseline methods, and recent large language models, including Llama 2 and GPT-4. The results show that by grouping tasks with similar PVI estimates, the joint learners yielded competitive results with fewer total parameters, with consistent performance across domains.
comment: main paper 12 pages, Appendix 7 pages, 1 figure, 18 tables
♻ ☆ ContextQFormer: A New Context Modeling Method for Multi-Turn Multi-Modal Conversations
Multi-modal large language models have demonstrated remarkable zero-shot abilities and powerful image-understanding capabilities. However, the existing open-source multi-modal models suffer from the weak capability of multi-turn interaction, especially for long contexts. To address the issue, we first introduce a context modeling module, termed ContextQFormer, which utilizes a memory block to enhance the presentation of contextual information. Furthermore, to facilitate further research, we carefully build a new multi-turn multi-modal dialogue dataset (TMDialog) for pre-training, instruction-tuning, and evaluation, which will be open-sourced lately. Compared with other multi-modal dialogue datasets, TMDialog contains longer conversations, which supports the research of multi-turn multi-modal dialogue. In addition, ContextQFormer is compared with three baselines on TMDialog and experimental results illustrate that ContextQFormer achieves an improvement of 2%-4% in available rate over baselines.
comment: 9 pages, 6 figures
♻ ☆ Secure Multifaceted-RAG for Enterprise: Hybrid Knowledge Retrieval with Security Filtering
Existing Retrieval-Augmented Generation (RAG) systems face challenges in enterprise settings due to limited retrieval scope and data security risks. When relevant internal documents are unavailable, the system struggles to generate accurate and complete responses. Additionally, using closed-source Large Language Models (LLMs) raises concerns about exposing proprietary information. To address these issues, we propose the Secure Multifaceted-RAG (SecMulti-RAG) framework, which retrieves not only from internal documents but also from two supplementary sources: pre-generated expert knowledge for anticipated queries and on-demand external LLM-generated knowledge. To mitigate security risks, we adopt a local open-source generator and selectively utilize external LLMs only when prompts are deemed safe by a filtering mechanism. This approach enhances completeness, prevents data leakage, and reduces costs. In our evaluation on a report generation task in the automotive industry, SecMulti-RAG significantly outperforms traditional RAG - achieving 79.3 to 91.9 percent win rates across correctness, richness, and helpfulness in LLM-based evaluation, and 56.3 to 70.4 percent in human evaluation. This highlights SecMulti-RAG as a practical and secure solution for enterprise RAG.
♻ ☆ ConTextual: Improving Clinical Text Summarization in LLMs with Context-preserving Token Filtering and Knowledge Graphs
Unstructured clinical data can serve as a unique and rich source of information that can meaningfully inform clinical practice. Extracting the most pertinent context from such data is critical for exploiting its true potential toward optimal and timely decision-making in patient care. While prior research has explored various methods for clinical text summarization, most prior studies either process all input tokens uniformly or rely on heuristic-based filters, which can overlook nuanced clinical cues and fail to prioritize information critical for decision-making. In this study, we propose Contextual, a novel framework that integrates a Context-Preserving Token Filtering method with a Domain-Specific Knowledge Graph (KG) for contextual augmentation. By preserving context-specific important tokens and enriching them with structured knowledge, ConTextual improves both linguistic coherence and clinical fidelity. Our extensive empirical evaluations on two public benchmark datasets demonstrate that ConTextual consistently outperforms other baselines. Our proposed approach highlights the complementary role of token-level filtering and structured retrieval in enhancing both linguistic and clinical integrity, as well as offering a scalable solution for improving precision in clinical text generation.
comment: Accepted for MLHC 2025
♻ ☆ CoDet-M4: Detecting Machine-Generated Code in Multi-Lingual, Multi-Generator and Multi-Domain Settings
Large language models (LLMs) have revolutionized code generation, automating programming with remarkable efficiency. However, these advancements challenge programming skills, ethics, and assessment integrity, making the detection of LLM-generated code essential for maintaining accountability and standards. While, there has been some research on this problem, it generally lacks domain coverage and robustness, and only covers a small number of programming languages. To this end, we propose a framework capable of distinguishing between human- and LLM-written code across multiple programming languages, code generators, and domains. We use a large-scale dataset from renowned platforms and LLM-based code generators, alongside applying rigorous data quality checks, feature engineering, and comparative analysis using evaluation of traditional machine learning models, pre-trained language models (PLMs), and LLMs for code detection. We perform an evaluation on out-of-domain scenarios, such as detecting the authorship and hybrid authorship of generated code and generalizing to unseen models, domains, and programming languages. Moreover, our extensive experiments show that our framework effectively distinguishes human- from LLM-written code and sets a new benchmark for this task.
♻ ☆ A Comparative Approach to Assessing Linguistic Creativity of Large Language Models and Humans
The following paper introduces a general linguistic creativity test for humans and Large Language Models (LLMs). The test consists of various tasks aimed at assessing their ability to generate new original words and phrases based on word formation processes (derivation and compounding) and on metaphorical language use. We administered the test to 24 humans and to an equal number of LLMs, and we automatically evaluated their answers using OCSAI tool for three criteria: Originality, Elaboration, and Flexibility. The results show that LLMs not only outperformed humans in all the assessed criteria, but did better in six out of the eight test tasks. We then computed the uniqueness of the individual answers, which showed some minor differences between humans and LLMs. Finally, we performed a short manual analysis of the dataset, which revealed that humans are more inclined towards E(extending)-creativity, while LLMs favor F(ixed)-creativity.
comment: Accepted for presentation at KES 2025. To appear in Procedia Computer Science (Elsevier)
♻ ☆ SWE-MERA: A Dynamic Benchmark for Agenticly Evaluating Large Language Models on Software Engineering Tasks
The rapid advancement of Large Language Models (LLMs) in software engineering has revealed critical limitations in existing benchmarks, particularly the widely used SWE-bench dataset. Recent studies have uncovered severe data contamination issues, e.g. SWE-bench reports 32.67% of successful patches involve direct solution leakage and 31.08% pass due to inadequate test cases. We introduce SWE-MERA, a dynamic, continuously updated benchmark designed to address these fundamental challenges through an automated collection of real-world GitHub issues and rigorous quality validation. Our approach implements a reliable pipeline that ensures quality while minimizing contamination risks, resulting in approximately 10,000 potential tasks with 300 samples currently available. Evaluation using the Aider coding agent demonstrates strong discriminative power in state-of-the-art models. We report performance across a dozen recent LLMs evaluated on tasks collected between September 2024 and June 2025.
♻ ☆ Task-Circuit Quantization: Leveraging Knowledge Localization and Interpretability for Compression
Post-training quantization (PTQ) reduces a model's memory footprint by mapping full precision weights into low bit weights without costly retraining, but can degrade its downstream performance especially in low 2- to 3-bit settings. We develop a new mixed-precision PTQ approach, Task-Circuit Quantization (TaCQ), that draws parallels to automated circuit discovery, directly conditioning the quantization process on specific weight circuits -- which we define as sets of weights associated with downstream task performance. These weights are kept as 16-bit weights, while others are quantized, maintaining performance while only adding a marginal memory cost. Specifically, TaCQ contrasts unquantized model weights with a uniformly-quantized model to estimate the expected change in weights due to quantization and uses gradient information to predict the resulting impact on task performance, allowing us to preserve task-specific weights. We compare TaCQ-based quantization to existing mixed-precision quantization methods when conditioning both on general-purpose and task-specific data. Across QA, math reasoning, and text-to-SQL tasks for both Llama-3 and Qwen2.5, we find that TaCQ outperforms baselines using the same calibration data and a lower weight budget, achieving major improvements in the 2 and 3-bit regime. With only 3.1 bits we are able to recover 96% of Llama-3-8B-Instruct's unquantized 16-bit MMLU performance, obtaining a 5.25% absolute improvement over SPQR. We also observe consistently large gains over existing methods in the 2-bit regime, with an average gain of 14.74% over the strongest baseline, SliM-LLM. Moreover, we observe a 7.20% gain without conditioning on specific tasks, showing TaCQ's ability to identify important weights is not limited to task-conditioned settings.
comment: COLM 2025 Camera Ready. Code: https://github.com/The-Inscrutable-X/TACQ
♻ ☆ Unified Triplet-Level Hallucination Evaluation for Large Vision-Language Models
Despite the outstanding performance in vision-language reasoning, Large Vision-Language Models (LVLMs) might generate hallucinated contents that do not exist in the given image. Most existing LVLM hallucination benchmarks are constrained to evaluate the object-related hallucinations. However, the potential hallucination on the relations between two objects, i.e., relation hallucination, still lacks investigation. To remedy that, we design a unified framework to measure the object and relation hallucination in LVLMs simultaneously. The core idea of our framework is to evaluate hallucinations via (object, relation, object) triplets extracted from LVLMs' responses, making it easily generalizable to different vision-language tasks. Based on our framework, we further introduce Tri-HE, a novel Triplet-level Hallucination Evaluation benchmark which can be used to study both object and relation hallucination at the same time. With comprehensive evaluations on Tri-HE, we observe that the relation hallucination issue is even more serious than object hallucination among existing LVLMs, highlighting a previously neglected problem towards reliable LVLMs. Moreover, based on our findings, we design a simple training-free approach that effectively mitigates hallucinations for LVLMs. Our dataset and code for the reproduction of our experiments are available publicly at https://github.com/wujunjie1998/Tri-HE.
comment: Accepted by TMLR 2025. Project Page: https://kaichen1998.github.io/projects/tri-he/
♻ ☆ SmartThinker: Learning to Compress and Preserve Reasoning by Step-Level Length Control
Large reasoning models (LRMs) have exhibited remarkable reasoning capabilities through inference-time scaling, but this progress has also introduced considerable redundancy and inefficiency into their reasoning processes, resulting in substantial computational waste. Previous work has attempted to mitigate this issue by penalizing the overall length of generated samples during reinforcement learning (RL), with the goal of encouraging a more concise chains of thought. However, we observe that such global length penalty often lead to excessive compression of critical reasoning steps while preserving unnecessary details in simpler ones, yielding a suboptimal trade-off between accuracy and efficiency. To address this issue, we propose SmartThinker, a two-stage learnable framework designed to enable fine-grained control over the length of reasoning chains based on the importance of each individual step. In the first stage, SmartThinker adapts a reasoning model to a short-form reasoning mode through rejection sampling combined with supervised fine-tuning (SFT). In the second stage, SmartThinker applies Step-Level Length Control Policy Optimization (SCPO) to refine the model output distribution, which increases the proportion of length allocated to critical steps while reducing redundancy in less important ones. SCPO consists of four core components: an online importance estimator, a step-level length control reward function, a step-level generalized advantage estimation (S-GAE) and a difficulty-adaptive clipping strategy. Working in concert, these components enable SCPO to implement differentiated length control across reasoning steps. Empirical results across multiple reasoning benchmarks and various backbone models demonstrate that SmartThinker significantly reduces redundant reasoning while achieving comparable or even superior performance to existing methods.
♻ ☆ MERA Code: A Unified Framework for Evaluating Code Generation Across Tasks
Advancements in LLMs have enhanced task automation in software engineering; however, current evaluations primarily focus on natural language tasks, overlooking code quality. Most benchmarks prioritize high-level reasoning over executable code and real-world performance, leaving gaps in understanding true capabilities and risks associated with these models in production. To address this issue, we propose MERA Code, a new addition to the MERA benchmark family, specifically focused on evaluating code for the latest code generation LLMs in Russian. This benchmark includes 11 evaluation tasks that span 8 programming languages. Our proposed evaluation methodology features a taxonomy that outlines the practical coding skills necessary for models to complete these tasks. The benchmark comprises an open-source codebase for users to conduct MERA assessments, a scoring system compatible with various programming environments, and a platform featuring a leaderboard and submission system. We evaluate open LLMs and frontier API models, analyzing their limitations in terms of practical coding tasks in non-English languages. We are publicly releasing MERA to guide future research, anticipate groundbreaking features in model development, and standardize evaluation procedures.
♻ ☆ Learning to Translate Ambiguous Terminology by Preference Optimization on Post-Edits
In real world translation scenarios, terminology is rarely one-to-one. Instead, multiple valid translations may appear in a terminology dictionary, but correctness of a translation depends on corporate style guides and context. This can be challenging for neural machine translation (NMT) systems. Luckily, in a corporate context, many examples of human post-edits of valid but incorrect terminology exist. The goal of this work is to learn how to disambiguate our terminology based on these corrections. Our approach is based on preference optimization, using the term post-edit as the knowledge to be preferred. While previous work had to rely on unambiguous translation dictionaries to set hard constraints during decoding, or to add soft constraints in the input, our framework requires neither one-to-one dictionaries nor human intervention at decoding time. We report results on English-German post-edited data and find that the optimal combination of supervised fine-tuning and preference optimization, with both term-specific and full sequence objectives, yields statistically significant improvements in term accuracy over a strong NMT baseline without significant losses in COMET score. Additionally, we release test sets from our post-edited data and terminology dictionary.
♻ ☆ OASIS: Order-Augmented Strategy for Improved Code Search
Code embeddings capture the semantic representations of code and are crucial for various code-related large language model (LLM) applications, such as code search. Previous training primarily relies on optimizing the InfoNCE loss by comparing positive natural language (NL)-code pairs with in-batch negatives. However, due to the sparse nature of code contexts, training solely by comparing the major differences between positive and negative pairs may fail to capture deeper semantic nuances. To address this issue, we propose a novel order-augmented strategy for improved code search (OASIS). It leverages order-based similarity labels to train models to capture subtle differences in similarity among negative pairs. Extensive benchmark evaluations demonstrate that our OASIS model significantly outperforms previous state-of-the-art models focusing solely on major positive-negative differences. It underscores the value of exploiting subtle differences among negative pairs with order labels for effective code embedding training.
♻ ☆ MEM1: Learning to Synergize Memory and Reasoning for Efficient Long-Horizon Agents
Modern language agents must operate over long-horizon, multi-turn interactions, where they retrieve external information, adapt to observations, and answer interdependent queries. Yet, most LLM systems rely on full-context prompting, appending all past turns regardless of their relevance. This leads to unbounded memory growth, increased computational costs, and degraded reasoning performance on out-of-distribution input lengths. We introduce MEM1, an end-to-end reinforcement learning framework that enables agents to operate with constant memory across long multi-turn tasks. At each turn, MEM1 updates a compact shared internal state that jointly supports memory consolidation and reasoning. This state integrates prior memory with new observations from the environment while strategically discarding irrelevant or redundant information. To support training in more realistic and compositional settings, we propose a simple yet effective and scalable approach to constructing multi-turn environments by composing existing datasets into arbitrarily complex task sequences. Experiments across three domains, including internal retrieval QA, open-domain web QA, and multi-turn web shopping, show that MEM1-7B improves performance by 3.5x while reducing memory usage by 3.7x compared to Qwen2.5-14B-Instruct on a 16-objective multi-hop QA task, and generalizes beyond the training horizon. Our results demonstrate the promise of reasoning-driven memory consolidation as a scalable alternative to existing solutions for training long-horizon interactive agents, where both efficiency and performance are optimized.
♻ ☆ Code2Logic: Game-Code-Driven Data Synthesis for Enhancing VLMs General Reasoning NeurIPS 2025
Visual-language Chain-of-Thought (CoT) data resources are relatively scarce compared to text-only counterparts, limiting the improvement of reasoning capabilities in Vision Language Models (VLMs). However, high-quality vision-language reasoning data is expensive and labor-intensive to annotate. To address this issue, we leverage a promising resource: game code, which naturally contains logical structures and state transition processes. Therefore, we propose Code2Logic, a novel game-code-driven approach for multimodal reasoning data synthesis. Our approach leverages Large Language Models (LLMs) to adapt game code, enabling automatic acquisition of reasoning processes and results through code execution. Using the Code2Logic approach, we developed the GameQA dataset to train and evaluate VLMs. GameQA is cost-effective and scalable, offers controllable difficulty gradation and is diverse with 30 games and 158 tasks. Surprisingly, despite training solely on game data, VLMs demonstrated out of domain generalization, specifically Qwen2.5-VL-7B improving performance by 2.33% across 7 diverse vision-language benchmarks. Our code, dataset and models are available at https://github.com/tongjingqi/Code2Logic.
comment: 63 pages, 23 figures, submitted to NeurIPS 2025
♻ ☆ IOPO: Empowering LLMs with Complex Instruction Following via Input-Output Preference Optimization ACL 2025
In the realm of large language models (LLMs), the ability of models to accurately follow instructions is paramount as more agents and applications leverage LLMs for construction, where the complexity of instructions are rapidly increasing. However, on the one hand, there is only a certain amount of complex instruction evaluation data; on the other hand, there are no dedicated algorithms to improve the ability to follow complex instructions. To this end, this paper introduces TRACE, a benchmark for improving and evaluating the complex instructionfollowing ability, which consists of 120K training data and 1K evaluation data. Furthermore, we propose IOPO (Input-Output Preference Optimization) alignment method which takes both input and output preference pairs into consideration, where LLMs not only rapidly align with response preferences but also meticulously explore the instruction preferences. Extensive experiments on both in-domain and outof-domain datasets confirm the effectiveness of IOPO, showing 8.15%, 2.18% improvements on in-domain data and 6.29%, 3.13% on outof-domain data compared to SFT and DPO respectively.
comment: ACL 2025
♻ ☆ On the Limitations of Large Language Models (LLMs): False Attribution
In this work, we introduce a new hallucination metric - Simple Hallucination Index (SHI) and provide insight into one important limitation of the parametric knowledge of large language models (LLMs), i.e. false attribution. The task of automatic author attribution for relatively small chunks of text is an important NLP task but can be challenging. We empirically evaluate the power of 3 open SotA LLMs in zero-shot setting (Gemma-7B, Mixtral 8x7B, and LLaMA-2-13B). We acquired the top 10 most popular books of a month, according to Project Gutenberg, divided each one into equal chunks of 400 words, and prompted each LLM to predict the author. We then randomly sampled 162 chunks per book for human evaluation, based on the error margin of 7% and a confidence level of 95%. The average results show that Mixtral 8x7B has the highest prediction accuracy, the lowest SHI, and a Pearson's correlation (r) of 0.724, 0.263, and -0.9996, respectively, followed by LLaMA-2-13B and Gemma-7B. However, Mixtral 8x7B suffers from high hallucinations for 3 books, rising as high as a SHI of 0.87 (in the range 0-1, where 1 is the worst). The strong negative correlation of accuracy and SHI, given by r, demonstrates the fidelity of the new hallucination metric, which may generalize to other tasks. We also show that prediction accuracies correlate positively with the frequencies of Wikipedia instances of the book titles instead of the downloads and we perform error analyses of predictions. We publicly release the annotated chunks of data and our codes to aid the reproducibility and evaluation of other models.
comment: This paper was accepted for presentation by Recent Advances in NLP (RANLP) 2025 conference
♻ ☆ Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.
comment: 72 pages, 17 figures
♻ ☆ A Logically Consistent Chain-of-Thought Approach for Stance Detection
Zero-shot stance detection (ZSSD) aims to detect stances toward unseen targets. Incorporating background knowledge to enhance transferability between seen and unseen targets constitutes the primary approach of ZSSD. However, these methods often struggle with a knowledge-task disconnect and lack logical consistency in their predictions. To address these issues, we introduce a novel approach named Logically Consistent Chain-of-Thought (LC-CoT) for ZSSD, which improves stance detection by ensuring relevant and logically sound knowledge extraction. LC-CoT employs a three-step process. Initially, it assesses whether supplementary external knowledge is necessary. Subsequently, it uses API calls to retrieve this knowledge, which can be processed by a separate LLM. Finally, a manual exemplar guides the LLM to infer stance categories, using an if-then logical structure to maintain relevance and logical coherence. This structured approach to eliciting background knowledge enhances the model's capability, outperforming traditional supervised methods without relying on labeled data.
♻ ☆ MAC-Tuning: LLM Multi-Compositional Problem Reasoning with Enhanced Knowledge Boundary Awareness
With the widespread application of large language models (LLMs), the issue of generating non-existing facts, known as hallucination, has garnered increasing attention. Previous research in enhancing LLM confidence estimation mainly focuses on the single problem setting. However, LLM awareness of its internal parameterized knowledge boundary under the more challenging multi-problem setting, which requires answering multiple problems accurately simultaneously, remains underexplored. To bridge this gap, we introduce a novel method, Multiple Answers and Confidence Stepwise Tuning (MAC-Tuning), that separates the learning of answer prediction and confidence estimation during fine-tuning on instruction data. Extensive experiments demonstrate that our method outperforms baselines by up to 25% in average precision.
♻ ☆ SEALGuard: Safeguarding the Multilingual Conversations in Southeast Asian Languages for LLM Software Systems
Safety alignment is critical for LLM-powered systems. While recent LLM-powered guardrail approaches such as LlamaGuard achieve high detection accuracy of unsafe inputs written in English (e.g., ``How to create a bomb?''), they struggle with multilingual unsafe inputs. This limitation leaves LLM systems vulnerable to unsafe and jailbreak prompts written in low-resource languages such as those in Southeast Asia. This paper introduces SEALGuard, a multilingual guardrail designed to improve the safety alignment across diverse languages. It aims to address the multilingual safety alignment gap of existing guardrails and ensure effective filtering of unsafe and jailbreak prompts in LLM-powered systems. We adapt a general-purpose multilingual language model into a multilingual guardrail using low-rank adaptation (LoRA). We construct SEALSBench, a large-scale multilingual safety alignment dataset containing over 260,000 prompts in ten languages, including safe, unsafe, and jailbreak cases. We evaluate SEALGuard against state-of-the-art guardrails such as LlamaGuard on this benchmark. Our findings show that multilingual unsafe and jailbreak prompts substantially degrade the performance of the state-of-the-art LlamaGuard, which experiences a drop in Defense Success Rate (DSR) by 9% and 18%, respectively, compared to its performance on English-only prompts. In contrast, SEALGuard outperforms existing guardrails in detecting multilingual unsafe and jailbreak prompts, improving DSR by 48% over LlamaGuard and achieving the best DSR, precision, and F1-score. Our ablation study further reveals the contributions of adaptation strategies and model size to the overall performance of SEALGuard. We release our pre-trained model and benchmark at https://github.com/awsm-research/SEALGuard to support further research.
♻ ☆ ReCode: Updating Code API Knowledge with Reinforcement Learning
Large Language Models (LLMs) exhibit remarkable code generation capabilities but falter when adapting to frequent updates in external library APIs. This critical limitation, stemming from reliance on outdated API knowledge from their training data, even with access to current documentation, impedes reliable code generation in dynamic environments. To tackle this issue, we propose ReCode (rule-based Reinforcement learning for Code Update), a novel framework that mimics human programmer adaptation to API changes. Specifically, we construct a dataset of approximately 2,000 data entries to train the LLMs to perform version migration based on updated information. Then, we introduce a modified string similarity metric for code evaluation as the reward for reinforcement learning. Our experiments demonstrate that ReCode substantially boosts LLMs' code generation performance in dynamic API scenarios, especially on the unseen CodeUpdateArena task. Crucially, compared to supervised fine-tuning, ReCode has less impact on LLMs' general code generation abilities. We apply ReCode on various LLMs and reinforcement learning algorithms (GRPO and DAPO), all achieving consistent improvements. Notably, after training, Qwen2.5-Coder-7B outperforms that of the 32B parameter code instruction-tuned model and the reasoning model with the same architecture. Code is available at https://github.com/zjunlp/ReCode.
comment: Work in progress
♻ ☆ MPO: An Efficient Post-Processing Framework for Mixing Diverse Preference Alignment ICML 2025
Reinforcement Learning from Human Feedback (RLHF) has shown promise in aligning large language models (LLMs). Yet its reliance on a singular reward model often overlooks the diversity of human preferences. Recent approaches address this limitation by leveraging multi-dimensional feedback to fine-tune corresponding reward models and train LLMs using reinforcement learning. However, the process is costly and unstable, especially given the competing and heterogeneous nature of human preferences. In this paper, we propose Mixing Preference Optimization (MPO), a post-processing framework for aggregating single-objective policies as an alternative to both multi-objective RLHF (MORLHF) and MaxMin-RLHF. MPO avoids alignment from scratch. Instead, it log-linearly combines existing policies into a unified one with the weight of each policy computed via a batch stochastic mirror descent. Empirical results demonstrate that MPO achieves balanced performance across diverse preferences, outperforming or matching existing models with significantly reduced computational costs.
comment: ICML 2025
♻ ☆ Critique-GRPO: Advancing LLM Reasoning with Natural Language and Numerical Feedback
Recent advances in reinforcement learning (RL) with numerical feedback, such as scalar rewards, have significantly enhanced the complex reasoning capabilities of large language models (LLMs). Despite this success, we identify three key challenges encountered by RL with solely numerical feedback: performance plateaus, limited effectiveness of spontaneous self-reflection, and persistent failures. We then demonstrate that RL-finetuned models, even after exhibiting performance plateaus, can generate correct refinements on persistently failed problems by leveraging natural language feedback in the form of critiques. Building on this insight, we propose Critique-GRPO, an online RL framework that integrates both natural language and numerical feedback for effective policy optimization. Critique-GRPO enables LLMs to learn from initial responses and critique-guided self-refinements simultaneously while maintaining exploration. Additionally, we employ a shaping function to amplify learning from correct, especially unfamiliar, refinements and penalize incorrect ones. Extensive experiments with Qwen2.5-7B-Base, Qwen2.5-Math-7B-Base, and Qwen3-8B demonstrate that Critique-GRPO consistently outperforms supervised learning and RL-based fine-tuning methods across eight challenging mathematical, STEM, and general reasoning tasks, improving average pass@1 scores by approximately 4.4% and 3.8% on Qwen2.5-7B-Base and Qwen3-8B, respectively. Notably, Critique-GRPO enables effective self-improvement through self-critiquing and weak-to-strong generalization, achieving consistent gains over GRPO, such as 16.7% and 10.0% pass@1 improvements on AIME 2024, respectively.
comment: 52 pages, updated with new experimental results and implementation details
♻ ☆ VIDEE: Visual and Interactive Decomposition, Execution, and Evaluation of Text Analytics with Intelligent Agents
Text analytics has traditionally required specialized knowledge in Natural Language Processing (NLP) or text analysis, which presents a barrier for entry-level analysts. Recent advances in large language models (LLMs) have changed the landscape of NLP by enabling more accessible and automated text analysis (e.g., topic detection, summarization, information extraction, etc.). We introduce VIDEE, a system that supports entry-level data analysts to conduct advanced text analytics with intelligent agents. VIDEE instantiates a human-agent collaroration workflow consisting of three stages: (1) Decomposition, which incorporates a human-in-the-loop Monte-Carlo Tree Search algorithm to support generative reasoning with human feedback, (2) Execution, which generates an executable text analytics pipeline, and (3) Evaluation, which integrates LLM-based evaluation and visualizations to support user validation of execution results. We conduct two quantitative experiments to evaluate VIDEE's effectiveness and analyze common agent errors. A user study involving participants with varying levels of NLP and text analytics experience -- from none to expert -- demonstrates the system's usability and reveals distinct user behavior patterns. The findings identify design implications for human-agent collaboration, validate the practical utility of VIDEE for non-expert users, and inform future improvements to intelligent text analytics systems.
♻ ☆ BEARCUBS: A benchmark for computer-using web agents
Modern web agents possess computer use abilities that allow them to interact with webpages by sending commands to a virtual keyboard and mouse. While such agents have considerable potential to assist human users with complex tasks, evaluating their capabilities in real-world settings poses a major challenge. To this end, we introduce BEARCUBS, a "small but mighty" benchmark of 111 information-seeking questions designed to evaluate a web agent's ability to search, browse, and identify factual information from the web. Unlike prior web agent benchmarks, solving BEARCUBS requires (1) accessing live web content rather than synthetic or simulated pages, which captures the unpredictability of real-world web interactions; and (2) performing a broad range of multimodal interactions (e.g., video understanding, 3D navigation) that cannot be bypassed via text-based workarounds. Each question in BEARCUBS has a corresponding short, unambiguous answer and a human-validated browsing trajectory, allowing for transparent evaluation of agent performance and strategies. A human study confirms that BEARCUBS questions are solvable but non-trivial (84.7% human accuracy), revealing domain knowledge gaps and overlooked details as common failure points. By contrast, state-of-the-art computer-using agents underperform, with the best-scoring system (OpenAI's Operator) reaching only 23.4% accuracy. These results highlight critical areas for improvement, including reliable source selection and more powerful multimodal capabilities. To facilitate future research, BEARCUBS will be updated periodically to replace invalid or contaminated questions, keeping the benchmark fresh for future generations of web agents.
comment: 16 pages
♻ ☆ Synthesizing Privacy-Preserving Text Data via Finetuning without Finetuning Billion-Scale LLMs
Synthetic data offers a promising path to train models while preserving data privacy. Differentially private (DP) finetuning of large language models (LLMs) as data generator is effective, but is impractical when computation resources are limited. Meanwhile, prompt-based methods such as private evolution depend heavily on the manual prompts, and ineffectively use private information in their iterative data selection process. To overcome these limitations, we propose CTCL (Data Synthesis with ConTrollability and CLustering), a novel framework for generating privacy-preserving synthetic data without extensive prompt engineering or billion-scale LLM finetuning. CTCL pretrains a lightweight 140M conditional generator and a clustering-based topic model on large-scale public data. To further adapt to the private domain, the generator is DP finetuned on private data for fine-grained textual information, while the topic model extracts a DP histogram representing distributional information. The DP generator then samples according to the DP histogram to synthesize a desired number of data examples. Evaluation across five diverse domains demonstrates the effectiveness of our framework, particularly in the strong privacy regime. Systematic ablation validates the design of each framework component and highlights the scalability of our approach.
comment: Code available at https://github.com/tanyuqian/synthetic-private-data
♻ ☆ GUI Test Migration via Abstraction and Concretization
GUI test migration aims to produce test cases with events and assertions to test specific functionalities of a target app. Existing migration approaches typically focus on the widget-mapping paradigm that maps widgets from source apps to target apps. However, since different apps may implement the same functionality in different ways, direct mapping may result in incomplete or buggy test cases, thus significantly impacting the effectiveness of testing target functionality and the practical applicability of migration approaches. In this paper, we propose a new migration paradigm (i.e., the abstraction-concretization paradigm) that first abstracts the test logic for the target functionality and then utilizes this logic to generate the concrete GUI test case. Furthermore, we introduce MACdroid, the first approach that migrates GUI test cases based on this paradigm. Specifically, we propose an abstraction technique that utilizes source test cases from source apps targeting the same functionality to extract a general test logic for that functionality. Then, we propose a concretization technique that utilizes the general test logic to guide an LLM in generating the corresponding GUI test case (including events and assertions) for the target app. We evaluate MACdroid on two widely-used datasets (including 31 apps, 34 functionalities, and 123 test cases). On the FrUITeR dataset, the test cases generated by MACdroid successfully test 64% of the target functionalities, improving the baselines by 191%. On the Lin dataset, MACdroid successfully tests 75% of the target functionalities, outperforming the baselines by 42%. These results underscore the effectiveness of MACdroid in GUI test migration.
comment: This paper has been accepted for publication in ACM Transactions on Software Engineering and Methodology (TOSEM) in 2025. The official publication link is: https://dl.acm.org/doi/10.1145/3726525
♻ ☆ Fairness Is Not Enough: Auditing Competence and Intersectional Bias in AI-powered Resume Screening
The increasing use of generative AI for resume screening is predicated on the assumption that it offers an unbiased alternative to biased human decision-making. However, this belief fails to address a critical question: are these AI systems fundamentally competent at the evaluative tasks they are meant to perform? This study investigates the question of competence through a two-part audit of eight major AI platforms. Experiment 1 confirmed complex, contextual racial and gender biases, with some models penalizing candidates merely for the presence of demographic signals. Experiment 2, which evaluated core competence, provided a critical insight: some models that appeared unbiased were, in fact, incapable of performing a substantive evaluation, relying instead on superficial keyword matching. This paper introduces the "Illusion of Neutrality" to describe this phenomenon, where an apparent lack of bias is merely a symptom of a model's inability to make meaningful judgments. This study recommends that organizations and regulators adopt a dual-validation framework, auditing AI hiring tools for both demographic bias and demonstrable competence to ensure they are both equitable and effective.
comment: 34 pages, 4 figures
♻ ☆ ActionStudio: A Lightweight Framework for Data and Training of Large Action Models
Large Action models are essential for enabling autonomous agents to perform complex tasks. However, training such models remains challenging due to the diversity of agent environments and the complexity of noisy agentic data. Existing infrastructure offers limited support for scalable, agent-specific fine-tuning and standardized agent data processing. We introduce ActionStudio, a lightweight and extensible data and training framework designed for large action models. ActionStudio unifies diverse agent trajectories using our proposed Unified Format 2.0, supports a range of training workflows with optimized multi-node distributed setup, and integrates robust preprocessing and real-time verification tools. ActionStudio demonstrates up to 9x higher throughput compared to existing agentic training frameworks, and our trained models yield top performances across public and realistic agent benchmarks. To support the broader research community, we open-source the ActionStudio framework and release actionstudio-98k, a curated dataset of 98k high-quality trajectories. Code: https://github.com/SalesforceAIResearch/xLAM.
comment: 16 pages; large action models; xLAM; ActionStudio
♻ ☆ Chain-of-Thought Prompting Obscures Hallucination Cues in Large Language Models: An Empirical Evaluation
Large Language Models (LLMs) often exhibit \textit{hallucinations}, generating factually incorrect or semantically irrelevant content in response to prompts. Chain-of-Thought (CoT) prompting can mitigate hallucinations by encouraging step-by-step reasoning, but its impact on hallucination detection remains underexplored. To bridge this gap, we conduct a systematic empirical evaluation. We begin with a pilot experiment, revealing that CoT reasoning significantly affects the LLM's internal states and token probability distributions. Building on this, we evaluate the impact of various CoT prompting methods on mainstream hallucination detection methods across both instruction-tuned and reasoning-oriented LLMs. Specifically, we examine three key dimensions: changes in hallucination score distributions, variations in detection accuracy, and shifts in detection confidence. Our findings show that while CoT prompting helps reduce hallucination frequency, it also tends to obscure critical signals used for detection, impairing the effectiveness of various detection methods. Our study highlights an overlooked trade-off in the use of reasoning. Code is publicly available at: https://anonymous.4open.science/r/cot-hallu-detect.
♻ ☆ Exploiting Adaptive Contextual Masking for Aspect-Based Sentiment Analysis PAKDD 2024
Aspect-Based Sentiment Analysis (ABSA) is a fine-grained linguistics problem that entails the extraction of multifaceted aspects, opinions, and sentiments from the given text. Both standalone and compound ABSA tasks have been extensively used in the literature to examine the nuanced information present in online reviews and social media posts. Current ABSA methods often rely on static hyperparameters for attention-masking mechanisms, which can struggle with context adaptation and may overlook the unique relevance of words in varied situations. This leads to challenges in accurately analyzing complex sentences containing multiple aspects with differing sentiments. In this work, we present adaptive masking methods that remove irrelevant tokens based on context to assist in Aspect Term Extraction and Aspect Sentiment Classification subtasks of ABSA. We show with our experiments that the proposed methods outperform the baseline methods in terms of accuracy and F1 scores on four benchmark online review datasets. Further, we show that the proposed methods can be extended with multiple adaptations and demonstrate a qualitative analysis of the proposed approach using sample text for aspect term extraction.
comment: 12 pages, 4 figures, Accepted at PAKDD 2024
Computer Vision and Pattern Recognition 15
☆ VideoITG: Multimodal Video Understanding with Instructed Temporal Grounding
Recent studies have revealed that selecting informative and relevant video frames can significantly improve the performance of Video Large Language Models (Video-LLMs). Current methods, such as reducing inter-frame redundancy, employing separate models for image-text relevance assessment, or utilizing temporal video grounding for event localization, substantially adopt unsupervised learning paradigms, whereas they struggle to address the complex scenarios in long video understanding. We propose Instructed Temporal Grounding for Videos (VideoITG), featuring customized frame sampling aligned with user instructions. The core of VideoITG is the VidThinker pipeline, an automated annotation framework that explicitly mimics the human annotation process. First, it generates detailed clip-level captions conditioned on the instruction; then, it retrieves relevant video segments through instruction-guided reasoning; finally, it performs fine-grained frame selection to pinpoint the most informative visual evidence. Leveraging VidThinker, we construct the VideoITG-40K dataset, containing 40K videos and 500K instructed temporal grounding annotations. We then design a plug-and-play VideoITG model, which takes advantage of visual language alignment and reasoning capabilities of Video-LLMs, for effective frame selection in a discriminative manner. Coupled with Video-LLMs, VideoITG achieves consistent performance improvements across multiple multimodal video understanding benchmarks, showing its superiority and great potentials for video understanding.
comment: Technical Report
☆ Hierarchical Rectified Flow Matching with Mini-Batch Couplings
Flow matching has emerged as a compelling generative modeling approach that is widely used across domains. To generate data via a flow matching model, an ordinary differential equation (ODE) is numerically solved via forward integration of the modeled velocity field. To better capture the multi-modality that is inherent in typical velocity fields, hierarchical flow matching was recently introduced. It uses a hierarchy of ODEs that are numerically integrated when generating data. This hierarchy of ODEs captures the multi-modal velocity distribution just like vanilla flow matching is capable of modeling a multi-modal data distribution. While this hierarchy enables to model multi-modal velocity distributions, the complexity of the modeled distribution remains identical across levels of the hierarchy. In this paper, we study how to gradually adjust the complexity of the distributions across different levels of the hierarchy via mini-batch couplings. We show the benefits of mini-batch couplings in hierarchical rectified flow matching via compelling results on synthetic and imaging data. Code is available at https://riccizz.github.io/HRF_coupling.
comment: Project Page: https://riccizz.github.io/HRF_coupling
☆ VisionThink: Smart and Efficient Vision Language Model via Reinforcement Learning
Recent advancements in vision-language models (VLMs) have improved performance by increasing the number of visual tokens, which are often significantly longer than text tokens. However, we observe that most real-world scenarios do not require such an extensive number of visual tokens. While the performance drops significantly in a small subset of OCR-related tasks, models still perform accurately in most other general VQA tasks with only 1/4 resolution. Therefore, we propose to dynamically process distinct samples with different resolutions, and present a new paradigm for visual token compression, namely, VisionThink. It starts with a downsampled image and smartly decides whether it is sufficient for problem solving. Otherwise, the model could output a special token to request the higher-resolution image. Compared to existing Efficient VLM methods that compress tokens using fixed pruning ratios or thresholds, VisionThink autonomously decides whether to compress tokens case by case. As a result, it demonstrates strong fine-grained visual understanding capability on OCR-related tasks, and meanwhile saves substantial visual tokens on simpler tasks. We adopt reinforcement learning and propose the LLM-as-Judge strategy to successfully apply RL to general VQA tasks. Moreover, we carefully design a reward function and penalty mechanism to achieve a stable and reasonable image resize call ratio. Extensive experiments demonstrate the superiority, efficiency, and effectiveness of our method. Our code is available at https://github.com/dvlab-research/VisionThink.
comment: Code and models are available at https://github.com/dvlab-research/VisionThink
☆ $π^3$: Scalable Permutation-Equivariant Visual Geometry Learning
We introduce $\pi^3$, a feed-forward neural network that offers a novel approach to visual geometry reconstruction, breaking the reliance on a conventional fixed reference view. Previous methods often anchor their reconstructions to a designated viewpoint, an inductive bias that can lead to instability and failures if the reference is suboptimal. In contrast, $\pi^3$ employs a fully permutation-equivariant architecture to predict affine-invariant camera poses and scale-invariant local point maps without any reference frames. This design makes our model inherently robust to input ordering and highly scalable. These advantages enable our simple and bias-free approach to achieve state-of-the-art performance on a wide range of tasks, including camera pose estimation, monocular/video depth estimation, and dense point map reconstruction. Code and models are publicly available.
comment: Project page: https://yyfz.github.io/pi3/
☆ Imbalance in Balance: Online Concept Balancing in Generation Models ICCV2025
In visual generation tasks, the responses and combinations of complex concepts often lack stability and are error-prone, which remains an under-explored area. In this paper, we attempt to explore the causal factors for poor concept responses through elaborately designed experiments. We also design a concept-wise equalization loss function (IMBA loss) to address this issue. Our proposed method is online, eliminating the need for offline dataset processing, and requires minimal code changes. In our newly proposed complex concept benchmark Inert-CompBench and two other public test sets, our method significantly enhances the concept response capability of baseline models and yields highly competitive results with only a few codes.
comment: Accepted by ICCV2025
☆ AutoPartGen: Autogressive 3D Part Generation and Discovery
We introduce AutoPartGen, a model that generates objects composed of 3D parts in an autoregressive manner. This model can take as input an image of an object, 2D masks of the object's parts, or an existing 3D object, and generate a corresponding compositional 3D reconstruction. Our approach builds upon 3DShape2VecSet, a recent latent 3D representation with powerful geometric expressiveness. We observe that this latent space exhibits strong compositional properties, making it particularly well-suited for part-based generation tasks. Specifically, AutoPartGen generates object parts autoregressively, predicting one part at a time while conditioning on previously generated parts and additional inputs, such as 2D images, masks, or 3D objects. This process continues until the model decides that all parts have been generated, thus determining automatically the type and number of parts. The resulting parts can be seamlessly assembled into coherent objects or scenes without requiring additional optimization. We evaluate both the overall 3D generation capabilities and the part-level generation quality of AutoPartGen, demonstrating that it achieves state-of-the-art performance in 3D part generation.
comment: Project page: https://silent-chen.github.io/AutoPartGen/
☆ Diffuman4D: 4D Consistent Human View Synthesis from Sparse-View Videos with Spatio-Temporal Diffusion Models
This paper addresses the challenge of high-fidelity view synthesis of humans with sparse-view videos as input. Previous methods solve the issue of insufficient observation by leveraging 4D diffusion models to generate videos at novel viewpoints. However, the generated videos from these models often lack spatio-temporal consistency, thus degrading view synthesis quality. In this paper, we propose a novel sliding iterative denoising process to enhance the spatio-temporal consistency of the 4D diffusion model. Specifically, we define a latent grid in which each latent encodes the image, camera pose, and human pose for a certain viewpoint and timestamp, then alternately denoising the latent grid along spatial and temporal dimensions with a sliding window, and finally decode the videos at target viewpoints from the corresponding denoised latents. Through the iterative sliding, information flows sufficiently across the latent grid, allowing the diffusion model to obtain a large receptive field and thus enhance the 4D consistency of the output, while making the GPU memory consumption affordable. The experiments on the DNA-Rendering and ActorsHQ datasets demonstrate that our method is able to synthesize high-quality and consistent novel-view videos and significantly outperforms the existing approaches. See our project page for interactive demos and video results: https://diffuman4d.github.io/ .
comment: Project page: https://diffuman4d.github.io/
☆ Taming Diffusion Transformer for Real-Time Mobile Video Generation
Diffusion Transformers (DiT) have shown strong performance in video generation tasks, but their high computational cost makes them impractical for resource-constrained devices like smartphones, and real-time generation is even more challenging. In this work, we propose a series of novel optimizations to significantly accelerate video generation and enable real-time performance on mobile platforms. First, we employ a highly compressed variational autoencoder (VAE) to reduce the dimensionality of the input data without sacrificing visual quality. Second, we introduce a KD-guided, sensitivity-aware tri-level pruning strategy to shrink the model size to suit mobile platform while preserving critical performance characteristics. Third, we develop an adversarial step distillation technique tailored for DiT, which allows us to reduce the number of inference steps to four. Combined, these optimizations enable our model to achieve over 10 frames per second (FPS) generation on an iPhone 16 Pro Max, demonstrating the feasibility of real-time, high-quality video generation on mobile devices.
comment: 9 pages, 4 figures, 5 tables
☆ SpectraLift: Physics-Guided Spectral-Inversion Network for Self-Supervised Hyperspectral Image Super-Resolution
High-spatial-resolution hyperspectral images (HSI) are essential for applications such as remote sensing and medical imaging, yet HSI sensors inherently trade spatial detail for spectral richness. Fusing high-spatial-resolution multispectral images (HR-MSI) with low-spatial-resolution hyperspectral images (LR-HSI) is a promising route to recover fine spatial structures without sacrificing spectral fidelity. Most state-of-the-art methods for HSI-MSI fusion demand point spread function (PSF) calibration or ground truth high resolution HSI (HR-HSI), both of which are impractical to obtain in real world settings. We present SpectraLift, a fully self-supervised framework that fuses LR-HSI and HR-MSI inputs using only the MSI's Spectral Response Function (SRF). SpectraLift trains a lightweight per-pixel multi-layer perceptron (MLP) network using ($i$)~a synthetic low-spatial-resolution multispectral image (LR-MSI) obtained by applying the SRF to the LR-HSI as input, ($ii$)~the LR-HSI as the output, and ($iii$)~an $\ell_1$ spectral reconstruction loss between the estimated and true LR-HSI as the optimization objective. At inference, SpectraLift uses the trained network to map the HR-MSI pixel-wise into a HR-HSI estimate. SpectraLift converges in minutes, is agnostic to spatial blur and resolution, and outperforms state-of-the-art methods on PSNR, SAM, SSIM, and RMSE benchmarks.
☆ A Real-Time System for Egocentric Hand-Object Interaction Detection in Industrial Domains
Hand-object interaction detection remains an open challenge in real-time applications, where intuitive user experiences depend on fast and accurate detection of interactions with surrounding objects. We propose an efficient approach for detecting hand-objects interactions from streaming egocentric vision that operates in real time. Our approach consists of an action recognition module and an object detection module for identifying active objects upon confirmed interaction. Our Mamba model with EfficientNetV2 as backbone for action recognition achieves 38.52% p-AP on the ENIGMA-51 benchmark at 30fps, while our fine-tuned YOLOWorld reaches 85.13% AP for hand and object. We implement our models in a cascaded architecture where the action recognition and object detection modules operate sequentially. When the action recognition predicts a contact state, it activates the object detection module, which in turn performs inference on the relevant frame to detect and classify the active object.
comment: 12 pages, 4 figures, In International Conference on Image Analysis and Processing
☆ Revisiting Reliability in the Reasoning-based Pose Estimation Benchmark
The reasoning-based pose estimation (RPE) benchmark has emerged as a widely adopted evaluation standard for pose-aware multimodal large language models (MLLMs). Despite its significance, we identified critical reproducibility and benchmark-quality issues that hinder fair and consistent quantitative evaluations. Most notably, the benchmark utilizes different image indices from those of the original 3DPW dataset, forcing researchers into tedious and error-prone manual matching processes to obtain accurate ground-truth (GT) annotations for quantitative metrics (\eg, MPJPE, PA-MPJPE). Furthermore, our analysis reveals several inherent benchmark-quality limitations, including significant image redundancy, scenario imbalance, overly simplistic poses, and ambiguous textual descriptions, collectively undermining reliable evaluations across diverse scenarios. To alleviate manual effort and enhance reproducibility, we carefully refined the GT annotations through meticulous visual matching and publicly release these refined annotations as an open-source resource, thereby promoting consistent quantitative evaluations and facilitating future advancements in human pose-aware multimodal reasoning.
comment: To be presented as a poster at MMFM 2025
☆ FashionPose: Text to Pose to Relight Image Generation for Personalized Fashion Visualization
Realistic and controllable garment visualization is critical for fashion e-commerce, where users expect personalized previews under diverse poses and lighting conditions. Existing methods often rely on predefined poses, limiting semantic flexibility and illumination adaptability. To address this, we introduce FashionPose, the first unified text-to-pose-to-relighting generation framework. Given a natural language description, our method first predicts a 2D human pose, then employs a diffusion model to generate high-fidelity person images, and finally applies a lightweight relighting module, all guided by the same textual input. By replacing explicit pose annotations with text-driven conditioning, FashionPose enables accurate pose alignment, faithful garment rendering, and flexible lighting control. Experiments demonstrate fine-grained pose synthesis and efficient, consistent relighting, providing a practical solution for personalized virtual fashion display.
☆ DiffClean: Diffusion-based Makeup Removal for Accurate Age Estimation
Accurate age verification can protect underage users from unauthorized access to online platforms and e-commerce sites that provide age-restricted services. However, accurate age estimation can be confounded by several factors, including facial makeup that can induce changes to alter perceived identity and age to fool both humans and machines. In this work, we propose DiffClean which erases makeup traces using a text-guided diffusion model to defend against makeup attacks. DiffClean improves age estimation (minor vs. adult accuracy by 4.8%) and face verification (TMR by 8.9% at FMR=0.01%) over competing baselines on digitally simulated and real makeup images.
♻ ☆ EgoVLA: Learning Vision-Language-Action Models from Egocentric Human Videos
Real robot data collection for imitation learning has led to significant advancements in robotic manipulation. However, the requirement for robot hardware in the process fundamentally constrains the scale of the data. In this paper, we explore training Vision-Language-Action (VLA) models using egocentric human videos. The benefit of using human videos is not only for their scale but more importantly for the richness of scenes and tasks. With a VLA trained on human video that predicts human wrist and hand actions, we can perform Inverse Kinematics and retargeting to convert the human actions to robot actions. We fine-tune the model using a few robot manipulation demonstrations to obtain the robot policy, namely EgoVLA. We propose a simulation benchmark called Ego Humanoid Manipulation Benchmark, where we design diverse bimanual manipulation tasks with demonstrations. We fine-tune and evaluate EgoVLA with Ego Humanoid Manipulation Benchmark and show significant improvements over baselines and ablate the importance of human data. Videos can be found on our website: https://rchalyang.github.io/EgoVLA
comment: More videos can be found on our website: https://rchalyang.github.io/EgoVLA
♻ ☆ SIDDA: SInkhorn Dynamic Domain Adaptation for Image Classification with Equivariant Neural Networks
Modern neural networks (NNs) often do not generalize well in the presence of a "covariate shift"; that is, in situations where the training and test data distributions differ, but the conditional distribution of classification labels remains unchanged. In such cases, NN generalization can be reduced to a problem of learning more domain-invariant features. Domain adaptation (DA) methods include a range of techniques aimed at achieving this; however, these methods have struggled with the need for extensive hyperparameter tuning, which then incurs significant computational costs. In this work, we introduce SIDDA, an out-of-the-box DA training algorithm built upon the Sinkhorn divergence, that can achieve effective domain alignment with minimal hyperparameter tuning and computational overhead. We demonstrate the efficacy of our method on multiple simulated and real datasets of varying complexity, including simple shapes, handwritten digits, and real astronomical observations. SIDDA is compatible with a variety of NN architectures, and it works particularly well in improving classification accuracy and model calibration when paired with equivariant neural networks (ENNs). We find that SIDDA enhances the generalization capabilities of NNs, achieving up to a $\approx40\%$ improvement in classification accuracy on unlabeled target data. We also study the efficacy of DA on ENNs with respect to the varying group orders of the dihedral group $D_N$, and find that the model performance improves as the degree of equivariance increases. Finally, we find that SIDDA enhances model calibration on both source and target data--achieving over an order of magnitude improvement in the ECE and Brier score. SIDDA's versatility, combined with its automated approach to domain alignment, has the potential to advance multi-dataset studies by enabling the development of highly generalizable models.
comment: 25 pages, 5 figures, 4 tables. code available at: https://github.com/deepskies/SIDDA
Information Retrieval 16
☆ SGCL: Unifying Self-Supervised and Supervised Learning for Graph Recommendation RecSys 2025
Recommender systems (RecSys) are essential for online platforms, providing personalized suggestions to users within a vast sea of information. Self-supervised graph learning seeks to harness high-order collaborative filtering signals through unsupervised augmentation on the user-item bipartite graph, primarily leveraging a multi-task learning framework that includes both supervised recommendation loss and self-supervised contrastive loss. However, this separate design introduces additional graph convolution processes and creates inconsistencies in gradient directions due to disparate losses, resulting in prolonged training times and sub-optimal performance. In this study, we introduce a unified framework of Supervised Graph Contrastive Learning for recommendation (SGCL) to address these issues. SGCL uniquely combines the training of recommendation and unsupervised contrastive losses into a cohesive supervised contrastive learning loss, aligning both tasks within a single optimization direction for exceptionally fast training. Extensive experiments on three real-world datasets show that SGCL outperforms state-of-the-art methods, achieving superior accuracy and efficiency.
comment: Accepted in RecSys 2025. arXiv admin note: substantial text overlap with arXiv:2404.15954
☆ Efficiently Constructing Sparse Navigable Graphs
Graph-based nearest neighbor search methods have seen a surge of popularity in recent years, offering state-of-the-art performance across a wide variety of applications. Central to these methods is the task of constructing a sparse navigable search graph for a given dataset endowed with a distance function. Unfortunately, doing so is computationally expensive, so heuristics are universally used in practice. In this work, we initiate the study of fast algorithms with provable guarantees for search graph construction. For a dataset with $n$ data points, the problem of constructing an optimally sparse navigable graph can be framed as $n$ separate but highly correlated minimum set cover instances. This yields a naive $O(n^3)$ time greedy algorithm that returns a navigable graph whose sparsity is at most $O(\log n)$ higher than optimal. We improve significantly on this baseline, taking advantage of correlation between the set cover instances to leverage techniques from streaming and sublinear-time set cover algorithms. Combined with problem-specific pre-processing techniques, we present an $\tilde{O}(n^2)$ time algorithm for constructing an $O(\log n)$-approximate sparsest navigable graph under any distance function. The runtime of our method is optimal up to logarithmic factors under the Strong Exponential Time Hypothesis via a reduction from Monochromatic Closest Pair. Moreover, we prove that, as with general set cover, obtaining better than an $O(\log n)$-approximation is NP-hard, despite the significant additional structure present in the navigable graph problem. Finally, we show that our techniques can also beat cubic time for the closely related and practically important problems of constructing $\alpha$-shortcut reachable and $\tau$-monotonic graphs, which are also used for nearest neighbor search. For such graphs, we obtain $\tilde{O}(n^{2.5})$ time or better algorithms.
Overview of the TalentCLEF 2025: Skill and Job Title Intelligence for Human Capital Management
Advances in natural language processing and large language models are driving a major transformation in Human Capital Management, with a growing interest in building smart systems based on language technologies for talent acquisition, upskilling strategies, and workforce planning. However, the adoption and progress of these technologies critically depend on the development of reliable and fair models, properly evaluated on public data and open benchmarks, which have so far been unavailable in this domain. To address this gap, we present TalentCLEF 2025, the first evaluation campaign focused on skill and job title intelligence. The lab consists of two tasks: Task A - Multilingual Job Title Matching, covering English, Spanish, German, and Chinese; and Task B - Job Title-Based Skill Prediction, in English. Both corpora were built from real job applications, carefully anonymized, and manually annotated to reflect the complexity and diversity of real-world labor market data, including linguistic variability and gender-marked expressions. The evaluations included monolingual and cross-lingual scenarios and covered the evaluation of gender bias. TalentCLEF attracted 76 registered teams with more than 280 submissions. Most systems relied on information retrieval techniques built with multilingual encoder-based models fine-tuned with contrastive learning, and several of them incorporated large language models for data augmentation or re-ranking. The results show that the training strategies have a larger effect than the size of the model alone. TalentCLEF provides the first public benchmark in this field and encourages the development of robust, fair, and transferable language technologies for the labor market.
☆ Automating Steering for Safe Multimodal Large Language Models
Recent progress in Multimodal Large Language Models (MLLMs) has unlocked powerful cross-modal reasoning abilities, but also raised new safety concerns, particularly when faced with adversarial multimodal inputs. To improve the safety of MLLMs during inference, we introduce a modular and adaptive inference-time intervention technology, AutoSteer, without requiring any fine-tuning of the underlying model. AutoSteer incorporates three core components: (1) a novel Safety Awareness Score (SAS) that automatically identifies the most safety-relevant distinctions among the model's internal layers; (2) an adaptive safety prober trained to estimate the likelihood of toxic outputs from intermediate representations; and (3) a lightweight Refusal Head that selectively intervenes to modulate generation when safety risks are detected. Experiments on LLaVA-OV and Chameleon across diverse safety-critical benchmarks demonstrate that AutoSteer significantly reduces the Attack Success Rate (ASR) for textual, visual, and cross-modal threats, while maintaining general abilities. These findings position AutoSteer as a practical, interpretable, and effective framework for safer deployment of multimodal AI systems.
comment: Working in progress. 22 pages (8+ for main); 25 figures; 1 table
☆ SemCSE: Semantic Contrastive Sentence Embeddings Using LLM-Generated Summaries For Scientific Abstracts
We introduce SemCSE, an unsupervised method for learning semantic embeddings of scientific texts. Building on recent advances in contrastive learning for text embeddings, our approach leverages LLM-generated summaries of scientific abstracts to train a model that positions semantically related summaries closer together in the embedding space. This resulting objective ensures that the model captures the true semantic content of a text, in contrast to traditional citation-based approaches that do not necessarily reflect semantic similarity. To validate this, we propose a novel benchmark designed to assess a model's ability to understand and encode the semantic content of scientific texts, demonstrating that our method enforces a stronger semantic separation within the embedding space. Additionally, we evaluate SemCSE on the comprehensive SciRepEval benchmark for scientific text embeddings, where it achieves state-of-the-art performance among models of its size, thus highlighting the benefits of a semantically focused training approach.
☆ Generative Multi-Target Cross-Domain Recommendation
Recently, there has been a surge of interest in Multi-Target Cross-Domain Recommendation (MTCDR), which aims to enhance recommendation performance across multiple domains simultaneously. Existing MTCDR methods primarily rely on domain-shared entities (\eg users or items) to fuse and transfer cross-domain knowledge, which may be unavailable in non-overlapped recommendation scenarios. Some studies model user preferences and item features as domain-sharable semantic representations, which can be utilized to tackle the MTCDR task. Nevertheless, they often require extensive auxiliary data for pre-training. Developing more effective solutions for MTCDR remains an important area for further exploration. Inspired by recent advancements in generative recommendation, this paper introduces GMC, a generative paradigm-based approach for multi-target cross-domain recommendation. The core idea of GMC is to leverage semantically quantized discrete item identifiers as a medium for integrating multi-domain knowledge within a unified generative model. GMC first employs an item tokenizer to generate domain-shared semantic identifiers for each item, and then formulates item recommendation as a next-token generation task by training a domain-unified sequence-to-sequence model. To further leverage the domain information to enhance performance, we incorporate a domain-aware contrastive loss into the semantic identifier learning, and perform domain-specific fine-tuning on the unified recommender. Extensive experiments on five public datasets demonstrate the effectiveness of GMC compared to a range of baseline methods.
☆ Machine-Readable Ads: Accessibility and Trust Patterns for AI Web Agents interacting with Online Advertisements
Autonomous multimodal language models are rapidly evolving into web agents that can browse, click, and purchase items on behalf of users, posing a threat to display advertising designed for human eyes. Yet little is known about how these agents interact with ads or which design principles ensure reliable engagement. To address this, we ran a controlled experiment using a faithful clone of the news site TT.com, seeded with diverse ads: static banners, GIFs, carousels, videos, cookie dialogues, and paywalls. We ran 300 initial trials plus follow-ups using the Document Object Model (DOM)-centric Browser Use framework with GPT-4o, Claude 3.7 Sonnet, Gemini 2.0 Flash, and the pixel-based OpenAI Operator, across 10 realistic user tasks. Our results show these agents display severe satisficing: they never scroll beyond two viewports and ignore purely visual calls to action, clicking banners only when semantic button overlays or off-screen text labels are present. Critically, when sweepstake participation required a purchase, GPT-4o and Claude 3.7 Sonnet subscribed in 100% of trials, and Gemini 2.0 Flash in 70%, revealing gaps in cost-benefit analysis. We identified five actionable design principles-semantic overlays, hidden labels, top-left placement, static frames, and dialogue replacement, that make human-centric creatives machine-detectable without harming user experience. We also evaluated agent trustworthiness through "behavior patterns" such as cookie consent handling and subscription choices, highlighting model-specific risk boundaries and the urgent need for robust trust evaluation frameworks in real-world advertising.
☆ Bridging the Gap: Leveraging Retrieval-Augmented Generation to Better Understand Public Concerns about Vaccines
Vaccine hesitancy threatens public health, leading to delayed or rejected vaccines. Social media is a vital source for understanding public concerns, and traditional methods like topic modelling often struggle to capture nuanced opinions. Though trained for query answering, large Language Models (LLMs) often miss current events and community concerns. Additionally, hallucinations in LLMs can compromise public health communication. To address these limitations, we developed a tool (VaxPulse Query Corner) using the Retrieval Augmented Generation technique. It addresses complex queries about public vaccine concerns on various online platforms, aiding public health administrators and stakeholders in understanding public concerns and implementing targeted interventions to boost vaccine confidence. Analysing 35,103 Shingrix social media posts, it achieved answer faithfulness (0.96) and relevance (0.94).
☆ PinFM: Foundation Model for User Activity Sequences at a Billion-scale Visual Discovery Platform RecSys 2025
User activity sequences have emerged as one of the most important signals in recommender systems. We present a foundational model, PinFM, for understanding user activity sequences across multiple applications at a billion-scale visual discovery platform. We pretrain a transformer model with 20B+ parameters using extensive user activity data, then fine-tune it for specific applications, efficiently coupling it with existing models. While this pretraining-and-fine-tuning approach has been popular in other domains, such as Vision and NLP, its application in industrial recommender systems presents numerous challenges. The foundational model must be scalable enough to score millions of items every second while meeting tight cost and latency constraints imposed by these systems. Additionally, it should capture the interactions between user activities and other features and handle new items that were not present during the pretraining stage. We developed innovative techniques to address these challenges. Our infrastructure and algorithmic optimizations, such as the Deduplicated Cross-Attention Transformer (DCAT), improved our throughput by 600% on Pinterest internal data. We demonstrate that PinFM can learn interactions between user sequences and candidate items by altering input sequences, leading to a 20% increase in engagement with new items. PinFM is now deployed to help improve the experience of more than a half billion users across various applications.
comment: RecSys 2025
♻ ☆ Imagine All The Relevance: Scenario-Profiled Indexing with Knowledge Expansion for Dense Retrieval
Existing dense retrieval models struggle with reasoning-intensive retrieval task as they fail to capture implicit relevance that requires reasoning beyond surface-level semantic information. To address these challenges, we propose Scenario-Profiled Indexing with Knowledge Expansion (SPIKE), a dense retrieval framework that explicitly indexes implicit relevance by decomposing documents into scenario-based retrieval units. SPIKE organizes documents into scenario, which encapsulates the reasoning process necessary to uncover implicit relationships between hypothetical information needs and document content. SPIKE constructs a scenario-augmented dataset using a powerful teacher large language model (LLM), then distills these reasoning capabilities into a smaller, efficient scenario generator. During inference, SPIKE incorporates scenario-level relevance alongside document-level relevance, enabling reasoning-aware retrieval. Extensive experiments demonstrate that SPIKE consistently enhances retrieval performance across various query types and dense retrievers. It also enhances the retrieval experience for users through scenario and offers valuable contextual information for LLMs in retrieval-augmented generation (RAG).
comment: Accepted to COLM 2025
♻ ☆ LLM-Driven Dual-Level Multi-Interest Modeling for Recommendation
Recently, much effort has been devoted to modeling users' multi-interests based on their behaviors or auxiliary signals. However, existing methods often rely on heuristic assumptions, e.g., co-occurring items indicate the same interest of users, failing to capture user multi-interests aligning with real-world scenarios. While large language models (LLMs) show significant potential for multi-interest analysis due to their extensive knowledge and powerful reasoning capabilities, two key challenges remain. First, the granularity of LLM-driven multi-interests is agnostic, possibly leading to overly fine or coarse interest grouping. Second, individual user analysis provides limited insights due to the data sparsity issue. In this paper, we propose an LLM-driven dual-level multi-interest modeling framework for more effective recommendation. At the user-individual level, we exploit LLMs to flexibly allocate items engaged by users into different semantic clusters, indicating their diverse and distinct interests. To alleviate the agnostic generation of LLMs, we adaptively assign these semantic clusters to users' collaborative multi-interests learned from global user-item interactions, allowing the granularity to be automatically adjusted according to the user's behaviors using an alignment module. To alleviate the limited insights derived from individual users' behaviors, at the user-crowd level, we propose aggregating user cliques into synthesized users with rich behaviors for more comprehensive LLM-driven multi-interest analysis. We formulate a max covering problem to ensure the compactness and representativeness of synthesized users' behaviors, and then conduct contrastive learning based on their LLM-driven multi-interests to disentangle item representations among different interests. Experiments on real-world datasets show the superiority of our approach against state-of-the-art methods.
comment: 10 pages, 5 figures
♻ ☆ OASIS: Order-Augmented Strategy for Improved Code Search
Code embeddings capture the semantic representations of code and are crucial for various code-related large language model (LLM) applications, such as code search. Previous training primarily relies on optimizing the InfoNCE loss by comparing positive natural language (NL)-code pairs with in-batch negatives. However, due to the sparse nature of code contexts, training solely by comparing the major differences between positive and negative pairs may fail to capture deeper semantic nuances. To address this issue, we propose a novel order-augmented strategy for improved code search (OASIS). It leverages order-based similarity labels to train models to capture subtle differences in similarity among negative pairs. Extensive benchmark evaluations demonstrate that our OASIS model significantly outperforms previous state-of-the-art models focusing solely on major positive-negative differences. It underscores the value of exploiting subtle differences among negative pairs with order labels for effective code embedding training.
♻ ☆ MEM1: Learning to Synergize Memory and Reasoning for Efficient Long-Horizon Agents
Modern language agents must operate over long-horizon, multi-turn interactions, where they retrieve external information, adapt to observations, and answer interdependent queries. Yet, most LLM systems rely on full-context prompting, appending all past turns regardless of their relevance. This leads to unbounded memory growth, increased computational costs, and degraded reasoning performance on out-of-distribution input lengths. We introduce MEM1, an end-to-end reinforcement learning framework that enables agents to operate with constant memory across long multi-turn tasks. At each turn, MEM1 updates a compact shared internal state that jointly supports memory consolidation and reasoning. This state integrates prior memory with new observations from the environment while strategically discarding irrelevant or redundant information. To support training in more realistic and compositional settings, we propose a simple yet effective and scalable approach to constructing multi-turn environments by composing existing datasets into arbitrarily complex task sequences. Experiments across three domains, including internal retrieval QA, open-domain web QA, and multi-turn web shopping, show that MEM1-7B improves performance by 3.5x while reducing memory usage by 3.7x compared to Qwen2.5-14B-Instruct on a 16-objective multi-hop QA task, and generalizes beyond the training horizon. Our results demonstrate the promise of reasoning-driven memory consolidation as a scalable alternative to existing solutions for training long-horizon interactive agents, where both efficiency and performance are optimized.
♻ ☆ LLM-Enhanced User-Item Interactions: Leveraging Edge Information for Optimized Recommendations
Graph recommendation methods, representing a connected interaction perspective, reformulate user-item interactions as graphs to leverage graph structure and topology to recommend and have proved practical effectiveness at scale. Large language models, representing a textual generative perspective, excel at modeling user languages, understanding behavioral contexts, capturing user-item semantic relationships, analyzing textual sentiments, and generating coherent and contextually relevant texts as recommendations. However, there is a gap between the connected graph perspective and the text generation perspective as the task formulations are different. A research question arises: how can we effectively integrate the two perspectives for more personalized recsys? To fill this gap, we propose to incorporate graph-edge information into LLMs via prompt and attention innovations. We reformulate recommendations as a probabilistic generative problem using prompts. We develop a framework to incorporate graph edge information from the prompt and attention mechanisms for graph-structured LLM recommendations. We develop a new prompt design that brings in both first-order and second-order graph relationships; we devise an improved LLM attention mechanism to embed direct the spatial and connectivity information of edges. Our evaluation of real-world datasets demonstrates the framework's ability to understand connectivity information in graph data and to improve the relevance and quality of recommendation results.
♻ ☆ ReCode: Updating Code API Knowledge with Reinforcement Learning
Large Language Models (LLMs) exhibit remarkable code generation capabilities but falter when adapting to frequent updates in external library APIs. This critical limitation, stemming from reliance on outdated API knowledge from their training data, even with access to current documentation, impedes reliable code generation in dynamic environments. To tackle this issue, we propose ReCode (rule-based Reinforcement learning for Code Update), a novel framework that mimics human programmer adaptation to API changes. Specifically, we construct a dataset of approximately 2,000 data entries to train the LLMs to perform version migration based on updated information. Then, we introduce a modified string similarity metric for code evaluation as the reward for reinforcement learning. Our experiments demonstrate that ReCode substantially boosts LLMs' code generation performance in dynamic API scenarios, especially on the unseen CodeUpdateArena task. Crucially, compared to supervised fine-tuning, ReCode has less impact on LLMs' general code generation abilities. We apply ReCode on various LLMs and reinforcement learning algorithms (GRPO and DAPO), all achieving consistent improvements. Notably, after training, Qwen2.5-Coder-7B outperforms that of the 32B parameter code instruction-tuned model and the reasoning model with the same architecture. Code is available at https://github.com/zjunlp/ReCode.
comment: Work in progress
♻ ☆ LLM-RecG: A Semantic Bias-Aware Framework for Zero-Shot Sequential Recommendation
Zero-shot cross-domain sequential recommendation (ZCDSR) enables predictions in unseen domains without additional training or fine-tuning, addressing the limitations of traditional models in sparse data environments. Recent advancements in large language models (LLMs) have significantly enhanced ZCDSR by facilitating cross-domain knowledge transfer through rich, pretrained representations. Despite this progress, domain semantic bias -- arising from differences in vocabulary and content focus between domains -- remains a persistent challenge, leading to misaligned item embeddings and reduced generalization across domains. To address this, we propose a novel semantic bias-aware framework that enhances LLM-based ZCDSR by improving cross-domain alignment at both the item and sequential levels. At the item level, we introduce a generalization loss that aligns the embeddings of items across domains (inter-domain compactness), while preserving the unique characteristics of each item within its own domain (intra-domain diversity). This ensures that item embeddings can be transferred effectively between domains without collapsing into overly generic or uniform representations. At the sequential level, we develop a method to transfer user behavioral patterns by clustering source domain user sequences and applying attention-based aggregation during target domain inference. We dynamically adapt user embeddings to unseen domains, enabling effective zero-shot recommendations without requiring target-domain interactions...
comment: 10 pages, Recsys'25 Spotlight Oral
Machine Learning 150
☆ Hierarchical Rectified Flow Matching with Mini-Batch Couplings
Flow matching has emerged as a compelling generative modeling approach that is widely used across domains. To generate data via a flow matching model, an ordinary differential equation (ODE) is numerically solved via forward integration of the modeled velocity field. To better capture the multi-modality that is inherent in typical velocity fields, hierarchical flow matching was recently introduced. It uses a hierarchy of ODEs that are numerically integrated when generating data. This hierarchy of ODEs captures the multi-modal velocity distribution just like vanilla flow matching is capable of modeling a multi-modal data distribution. While this hierarchy enables to model multi-modal velocity distributions, the complexity of the modeled distribution remains identical across levels of the hierarchy. In this paper, we study how to gradually adjust the complexity of the distributions across different levels of the hierarchy via mini-batch couplings. We show the benefits of mini-batch couplings in hierarchical rectified flow matching via compelling results on synthetic and imaging data. Code is available at https://riccizz.github.io/HRF_coupling.
comment: Project Page: https://riccizz.github.io/HRF_coupling
☆ VisionThink: Smart and Efficient Vision Language Model via Reinforcement Learning
Recent advancements in vision-language models (VLMs) have improved performance by increasing the number of visual tokens, which are often significantly longer than text tokens. However, we observe that most real-world scenarios do not require such an extensive number of visual tokens. While the performance drops significantly in a small subset of OCR-related tasks, models still perform accurately in most other general VQA tasks with only 1/4 resolution. Therefore, we propose to dynamically process distinct samples with different resolutions, and present a new paradigm for visual token compression, namely, VisionThink. It starts with a downsampled image and smartly decides whether it is sufficient for problem solving. Otherwise, the model could output a special token to request the higher-resolution image. Compared to existing Efficient VLM methods that compress tokens using fixed pruning ratios or thresholds, VisionThink autonomously decides whether to compress tokens case by case. As a result, it demonstrates strong fine-grained visual understanding capability on OCR-related tasks, and meanwhile saves substantial visual tokens on simpler tasks. We adopt reinforcement learning and propose the LLM-as-Judge strategy to successfully apply RL to general VQA tasks. Moreover, we carefully design a reward function and penalty mechanism to achieve a stable and reasonable image resize call ratio. Extensive experiments demonstrate the superiority, efficiency, and effectiveness of our method. Our code is available at https://github.com/dvlab-research/VisionThink.
comment: Code and models are available at https://github.com/dvlab-research/VisionThink
☆ Latent Policy Steering with Embodiment-Agnostic Pretrained World Models
Learning visuomotor policies via imitation has proven effective across a wide range of robotic domains. However, the performance of these policies is heavily dependent on the number of training demonstrations, which requires expensive data collection in the real world. In this work, we aim to reduce data collection efforts when learning visuomotor robot policies by leveraging existing or cost-effective data from a wide range of embodiments, such as public robot datasets and the datasets of humans playing with objects (human data from play). Our approach leverages two key insights. First, we use optic flow as an embodiment-agnostic action representation to train a World Model (WM) across multi-embodiment datasets, and finetune it on a small amount of robot data from the target embodiment. Second, we develop a method, Latent Policy Steering (LPS), to improve the output of a behavior-cloned policy by searching in the latent space of the WM for better action sequences. In real world experiments, we observe significant improvements in the performance of policies trained with a small amount of data (over 50% relative improvement with 30 demonstrations and over 20% relative improvement with 50 demonstrations) by combining the policy with a WM pretrained on two thousand episodes sampled from the existing Open X-embodiment dataset across different robots or a cost-effective human dataset from play.
☆ Training Transformers with Enforced Lipschitz Constants
Neural networks are often highly sensitive to input and weight perturbations. This sensitivity has been linked to pathologies such as vulnerability to adversarial examples, divergent training, and overfitting. To combat these problems, past research has looked at building neural networks entirely from Lipschitz components. However, these techniques have not matured to the point where researchers have trained a modern architecture such as a transformer with a Lipschitz certificate enforced beyond initialization. To explore this gap, we begin by developing and benchmarking novel, computationally-efficient tools for maintaining norm-constrained weight matrices. Applying these tools, we are able to train transformer models with Lipschitz bounds enforced throughout training. We find that optimizer dynamics matter: switching from AdamW to Muon improves standard methods -- weight decay and spectral normalization -- allowing models to reach equal performance with a lower Lipschitz bound. Inspired by Muon's update having a fixed spectral norm, we co-design a weight constraint method that improves the Lipschitz vs. performance tradeoff on MLPs and 2M parameter transformers. Our 2-Lipschitz transformer on Shakespeare text reaches validation accuracy 60%. Scaling to 145M parameters, our 10-Lipschitz transformer reaches 21% accuracy on internet text. However, to match the NanoGPT baseline validation accuracy of 39.4%, our Lipschitz upper bound increases to 10^264. Nonetheless, our Lipschitz transformers train without stability measures such as layer norm, QK norm, and logit tanh softcapping.
☆ GeoReg: Weight-Constrained Few-Shot Regression for Socio-Economic Estimation using LLM
Socio-economic indicators like regional GDP, population, and education levels, are crucial to shaping policy decisions and fostering sustainable development. This research introduces GeoReg a regression model that integrates diverse data sources, including satellite imagery and web-based geospatial information, to estimate these indicators even for data-scarce regions such as developing countries. Our approach leverages the prior knowledge of large language model (LLM) to address the scarcity of labeled data, with the LLM functioning as a data engineer by extracting informative features to enable effective estimation in few-shot settings. Specifically, our model obtains contextual relationships between data features and the target indicator, categorizing their correlations as positive, negative, mixed, or irrelevant. These features are then fed into the linear estimator with tailored weight constraints for each category. To capture nonlinear patterns, the model also identifies meaningful feature interactions and integrates them, along with nonlinear transformations. Experiments across three countries at different stages of development demonstrate that our model outperforms baselines in estimating socio-economic indicators, even for low-income countries with limited data availability.
comment: 15 pages, 13 figures, 7 tables
☆ Boosting Team Modeling through Tempo-Relational Representation Learning
Team modeling remains a fundamental challenge at the intersection of Artificial Intelligence and the Social Sciences. Social Science research emphasizes the need to jointly model dynamics and relations, while practical applications demand unified models capable of inferring multiple team constructs simultaneously, providing interpretable insights and actionable recommendations to enhance team performance. However, existing works do not meet these practical demands. To bridge this gap, we present TRENN, a novel tempo-relational architecture that integrates: (i) an automatic temporal graph extractor, (ii) a tempo-relational encoder, (iii) a decoder for team construct prediction, and (iv) two complementary explainability modules. TRENN jointly captures relational and temporal team dynamics, providing a solid foundation for MT-TRENN, which extends TReNN by replacing the decoder with a multi-task head, enabling the model to learn shared Social Embeddings and simultaneously predict multiple team constructs, including Emergent Leadership, Leadership Style, and Teamwork components. Experimental results demonstrate that our approach significantly outperforms approaches that rely exclusively on temporal or relational information. Additionally, experimental evaluation has shown that the explainability modules integrated in MT-TRENN yield interpretable insights and actionable suggestions to support team improvement. These capabilities make our approach particularly well-suited for Human-Centered AI applications, such as intelligent decision-support systems in high-stakes collaborative environments.
☆ Optimal Empirical Risk Minimization under Temporal Distribution Shifts
Temporal distribution shifts pose a key challenge for machine learning models trained and deployed in dynamically evolving environments. This paper introduces RIDER (RIsk minimization under Dynamically Evolving Regimes) which derives optimally-weighted empirical risk minimization procedures under temporal distribution shifts. Our approach is theoretically grounded in the random distribution shift model, where random shifts arise as a superposition of numerous unpredictable changes in the data-generating process. We show that common weighting schemes, such as pooling all data, exponentially weighting data, and using only the most recent data, emerge naturally as special cases in our framework. We demonstrate that RIDER consistently improves out-of-sample predictive performance when applied as a fine-tuning step on the Yearbook dataset, across a range of benchmark methods in Wild-Time. Moreover, we show that RIDER outperforms standard weighting strategies in two other real-world tasks: predicting stock market volatility and forecasting ride durations in NYC taxi data.
☆ Stochastic Weakly Convex Optimization Under Heavy-Tailed Noises
An increasing number of studies have focused on stochastic first-order methods (SFOMs) under heavy-tailed gradient noises, which have been observed in the training of practical deep learning models. In this paper, we focus on two types of gradient noises: one is sub-Weibull noise, and the other is noise under the assumption that it has a bounded $p$-th central moment ($p$-BCM) with $p\in (1, 2]$. The latter is more challenging due to the occurrence of infinite variance when $p\in (1, 2)$. Under these two gradient noise assumptions, the in-expectation and high-probability convergence of SFOMs have been extensively studied in the contexts of convex optimization and standard smooth optimization. However, for weakly convex objectives-a class that includes all Lipschitz-continuous convex objectives and smooth objectives-our understanding of the in-expectation and high-probability convergence of SFOMs under these two types of noises remains incomplete. We investigate the high-probability convergence of the vanilla stochastic subgradient descent (SsGD) method under sub-Weibull noises, as well as the high-probability and in-expectation convergence of clipped SsGD under the $p$-BCM noises. Both analyses are conducted in the context of weakly convex optimization. For weakly convex objectives that may be non-convex and non-smooth, our results demonstrate that the theoretical dependence of vanilla SsGD on the failure probability and number of iterations under sub-Weibull noises does not degrade compared to the case of smooth objectives. Under $p$-BCM noises, our findings indicate that the non-smoothness and non-convexity of weakly convex objectives do not impact the theoretical dependence of clipped SGD on the failure probability relative to the smooth case; however, the sample complexity we derived is worse than a well-known lower bound for smooth optimization.
☆ Evaluating Reinforcement Learning Algorithms for Navigation in Simulated Robotic Quadrupeds: A Comparative Study Inspired by Guide Dog Behaviour
Robots are increasingly integrated across industries, particularly in healthcare. However, many valuable applications for quadrupedal robots remain overlooked. This research explores the effectiveness of three reinforcement learning algorithms in training a simulated quadruped robot for autonomous navigation and obstacle avoidance. The goal is to develop a robotic guide dog simulation capable of path following and obstacle avoidance, with long-term potential for real-world assistance to guide dogs and visually impaired individuals. It also seeks to expand research into medical 'pets', including robotic guide and alert dogs. A comparative analysis of thirteen related research papers shaped key evaluation criteria, including collision detection, pathfinding algorithms, sensor usage, robot type, and simulation platforms. The study focuses on sensor inputs, collision frequency, reward signals, and learning progression to determine which algorithm best supports robotic navigation in complex environments. Custom-made environments were used to ensure fair evaluation of all three algorithms under controlled conditions, allowing consistent data collection. Results show that Proximal Policy Optimization (PPO) outperformed Deep Q-Network (DQN) and Q-learning across all metrics, particularly in average and median steps to goal per episode. By analysing these results, this study contributes to robotic navigation, AI and medical robotics, offering insights into the feasibility of AI-driven quadruped mobility and its role in assistive robotics.
☆ Merge Kernel for Bayesian Optimization on Permutation Space AAAI-26
Bayesian Optimization (BO) algorithm is a standard tool for black-box optimization problems. The current state-of-the-art BO approach for permutation spaces relies on the Mallows kernel-an $\Omega(n^2)$ representation that explicitly enumerates every pairwise comparison. Inspired by the close relationship between the Mallows kernel and pairwise comparison, we propose a novel framework for generating kernel functions on permutation space based on sorting algorithms. Within this framework, the Mallows kernel can be viewed as a special instance derived from bubble sort. Further, we introduce the \textbf{Merge Kernel} constructed from merge sort, which replaces the quadratic complexity with $\Theta(n\log n)$ to achieve the lowest possible complexity. The resulting feature vector is significantly shorter, can be computed in linearithmic time, yet still efficiently captures meaningful permutation distances. To boost robustness and right-invariance without sacrificing compactness, we further incorporate three lightweight, task-agnostic descriptors: (1) a shift histogram, which aggregates absolute element displacements and supplies a global misplacement signal; (2) a split-pair line, which encodes selected long-range comparisons by aligning elements across the two halves of the whole permutation; and (3) sliding-window motifs, which summarize local order patterns that influence near-neighbor objectives. Our empirical evaluation demonstrates that the proposed kernel consistently outperforms the state-of-the-art Mallows kernel across various permutation optimization benchmarks. Results confirm that the Merge Kernel provides a more compact yet more effective solution for Bayesian optimization in permutation space.
comment: 8 pages, submitted to AAAI-26
☆ Automating Steering for Safe Multimodal Large Language Models
Recent progress in Multimodal Large Language Models (MLLMs) has unlocked powerful cross-modal reasoning abilities, but also raised new safety concerns, particularly when faced with adversarial multimodal inputs. To improve the safety of MLLMs during inference, we introduce a modular and adaptive inference-time intervention technology, AutoSteer, without requiring any fine-tuning of the underlying model. AutoSteer incorporates three core components: (1) a novel Safety Awareness Score (SAS) that automatically identifies the most safety-relevant distinctions among the model's internal layers; (2) an adaptive safety prober trained to estimate the likelihood of toxic outputs from intermediate representations; and (3) a lightweight Refusal Head that selectively intervenes to modulate generation when safety risks are detected. Experiments on LLaVA-OV and Chameleon across diverse safety-critical benchmarks demonstrate that AutoSteer significantly reduces the Attack Success Rate (ASR) for textual, visual, and cross-modal threats, while maintaining general abilities. These findings position AutoSteer as a practical, interpretable, and effective framework for safer deployment of multimodal AI systems.
comment: Working in progress. 22 pages (8+ for main); 25 figures; 1 table
☆ Leveraging Asynchronous Cross-border Market Data for Improved Day-Ahead Electricity Price Forecasting in European Markets
Accurate short-term electricity price forecasting is crucial for strategically scheduling demand and generation bids in day-ahead markets. While data-driven techniques have shown considerable prowess in achieving high forecast accuracy in recent years, they rely heavily on the quality of input covariates. In this paper, we investigate whether asynchronously published prices as a result of differing gate closure times (GCTs) in some bidding zones can improve forecasting accuracy in other markets with later GCTs. Using a state-of-the-art ensemble of models, we show significant improvements of 22% and 9% in forecast accuracy in the Belgian (BE) and Swedish bidding zones (SE3) respectively, when including price data from interconnected markets with earlier GCT (Germany-Luxembourg, Austria, and Switzerland). This improvement holds for both general as well as extreme market conditions. Our analysis also yields further important insights: frequent model recalibration is necessary for maximum accuracy but comes at substantial additional computational costs, and using data from more markets does not always lead to better performance - a fact we delve deeper into with interpretability analysis of the forecast models. Overall, these findings provide valuable guidance for market participants and decision-makers aiming to optimize bidding strategies within increasingly interconnected and volatile European energy markets.
comment: Both Maria Margarida Mascarenhas and Jilles De Blauwe contributed equally to the paper
☆ The carbon cost of materials discovery: Can machine learning really accelerate the discovery of new photovoltaics?
Computational screening has become a powerful complement to experimental efforts in the discovery of high-performance photovoltaic (PV) materials. Most workflows rely on density functional theory (DFT) to estimate electronic and optical properties relevant to solar energy conversion. Although more efficient than laboratory-based methods, DFT calculations still entail substantial computational and environmental costs. Machine learning (ML) models have recently gained attention as surrogates for DFT, offering drastic reductions in resource use with competitive predictive performance. In this study, we reproduce a canonical DFT-based workflow to estimate the maximum efficiency limit and progressively replace its components with ML surrogates. By quantifying the CO$_2$ emissions associated with each computational strategy, we evaluate the trade-offs between predictive efficacy and environmental cost. Our results reveal multiple hybrid ML/DFT strategies that optimize different points along the accuracy--emissions front. We find that direct prediction of scalar quantities, such as maximum efficiency, is significantly more tractable than using predicted absorption spectra as an intermediate step. Interestingly, ML models trained on DFT data can outperform DFT workflows using alternative exchange--correlation functionals in screening applications, highlighting the consistency and utility of data-driven approaches. We also assess strategies to improve ML-driven screening through expanded datasets and improved model architectures tailored to PV-relevant features. This work provides a quantitative framework for building low-emission, high-throughput discovery pipelines.
☆ Computational-Statistical Tradeoffs from NP-hardness
A central question in computer science and statistics is whether efficient algorithms can achieve the information-theoretic limits of statistical problems. Many computational-statistical tradeoffs have been shown under average-case assumptions, but since statistical problems are average-case in nature, it has been a challenge to base them on standard worst-case assumptions. In PAC learning where such tradeoffs were first studied, the question is whether computational efficiency can come at the cost of using more samples than information-theoretically necessary. We base such tradeoffs on $\mathsf{NP}$-hardness and obtain: $\circ$ Sharp computational-statistical tradeoffs assuming $\mathsf{NP}$ requires exponential time: For every polynomial $p(n)$, there is an $n$-variate class $C$ with VC dimension $1$ such that the sample complexity of time-efficiently learning $C$ is $\Theta(p(n))$. $\circ$ A characterization of $\mathsf{RP}$ vs. $\mathsf{NP}$ in terms of learning: $\mathsf{RP} = \mathsf{NP}$ iff every $\mathsf{NP}$-enumerable class is learnable with $O(\mathrm{VCdim}(C))$ samples in polynomial time. The forward implication has been known since (Pitt and Valiant, 1988); we prove the reverse implication. Notably, all our lower bounds hold against improper learners. These are the first $\mathsf{NP}$-hardness results for improperly learning a subclass of polynomial-size circuits, circumventing formal barriers of Applebaum, Barak, and Xiao (2008).
comment: To appear at FOCS 2025
☆ MoTM: Towards a Foundation Model for Time Series Imputation based on Continuous Modeling ALT
Recent years have witnessed a growing interest for time series foundation models, with a strong emphasis on the forecasting task. Yet, the crucial task of out-of-domain imputation of missing values remains largely underexplored. We propose a first step to fill this gap by leveraging implicit neural representations (INRs). INRs model time series as continuous functions and naturally handle various missing data scenarios and sampling rates. While they have shown strong performance within specific distributions, they struggle under distribution shifts. To address this, we introduce MoTM (Mixture of Timeflow Models), a step toward a foundation model for time series imputation. Building on the idea that a new time series is a mixture of previously seen patterns, MoTM combines a basis of INRs, each trained independently on a distinct family of time series, with a ridge regressor that adapts to the observed context at inference. We demonstrate robust in-domain and out-of-domain generalization across diverse imputation scenarios (e.g., block and pointwise missingness, variable sampling rates), paving the way for adaptable foundation imputation models.
comment: 10th Workshop on Advanced Analytics and Learning on Temporal Data (AALTD), ECML 2025
☆ Relation-Aware Slicing in Cross-Domain Alignment
The Sliced Gromov-Wasserstein (SGW) distance, aiming to relieve the computational cost of solving a non-convex quadratic program that is the Gromov-Wasserstein distance, utilizes projecting directions sampled uniformly from unit hyperspheres. This slicing mechanism incurs unnecessary computational costs due to uninformative directions, which also affects the representative power of the distance. However, finding a more appropriate distribution over the projecting directions (slicing distribution) is often an optimization problem in itself that comes with its own computational cost. In addition, with more intricate distributions, the sampling itself may be expensive. As a remedy, we propose an optimization-free slicing distribution that provides fast sampling for the Monte Carlo approximation. We do so by introducing the Relation-Aware Projecting Direction (RAPD), effectively capturing the pairwise association of each of two pairs of random vectors, each following their ambient law. This enables us to derive the Relation-Aware Slicing Distribution (RASD), a location-scale law corresponding to sampled RAPDs. Finally, we introduce the RASGW distance and its variants, e.g., IWRASGW (Importance Weighted RASGW), which overcome the shortcomings experienced by SGW. We theoretically analyze its properties and substantiate its empirical prowess using extensive experiments on various alignment tasks.
☆ GradNetOT: Learning Optimal Transport Maps with GradNets
Monotone gradient functions play a central role in solving the Monge formulation of the optimal transport problem, which arises in modern applications ranging from fluid dynamics to robot swarm control. When the transport cost is the squared Euclidean distance, Brenier's theorem guarantees that the unique optimal map is the gradient of a convex function, namely a monotone gradient map, and it satisfies a Monge-Amp\`ere equation. In [arXiv:2301.10862] [arXiv:2404.07361], we proposed Monotone Gradient Networks (mGradNets), neural networks that directly parameterize the space of monotone gradient maps. In this work, we leverage mGradNets to directly learn the optimal transport mapping by minimizing a training loss function defined using the Monge-Amp\`ere equation. We empirically show that the structural bias of mGradNets facilitates the learning of optimal transport maps and employ our method for a robot swarm control problem.
☆ Spectral Bellman Method: Unifying Representation and Exploration in RL
The effect of representation has been demonstrated in reinforcement learning, from both theoretical and empirical successes. However, the existing representation learning mainly induced from model learning aspects, misaligning with our RL tasks. This work introduces Spectral Bellman Representation, a novel framework derived from the Inherent Bellman Error (IBE) condition, which aligns with the fundamental structure of Bellman updates across a space of possible value functions, therefore, directly towards value-based RL. Our key insight is the discovery of a fundamental spectral relationship: under the zero-IBE condition, the transformation of a distribution of value functions by the Bellman operator is intrinsically linked to the feature covariance structure. This spectral connection yields a new, theoretically-grounded objective for learning state-action features that inherently capture this Bellman-aligned covariance. Our method requires a simple modification to existing algorithms. We demonstrate that our learned representations enable structured exploration, by aligning feature covariance with Bellman dynamics, and improve overall performance, particularly in challenging hard-exploration and long-horizon credit assignment tasks. Our framework naturally extends to powerful multi-step Bellman operators, further broadening its impact. Spectral Bellman Representation offers a principled and effective path toward learning more powerful and structurally sound representations for value-based reinforcement learning.
☆ SHIELD: A Secure and Highly Enhanced Integrated Learning for Robust Deepfake Detection against Adversarial Attacks
Audio plays a crucial role in applications like speaker verification, voice-enabled smart devices, and audio conferencing. However, audio manipulations, such as deepfakes, pose significant risks by enabling the spread of misinformation. Our empirical analysis reveals that existing methods for detecting deepfake audio are often vulnerable to anti-forensic (AF) attacks, particularly those attacked using generative adversarial networks. In this article, we propose a novel collaborative learning method called SHIELD to defend against generative AF attacks. To expose AF signatures, we integrate an auxiliary generative model, called the defense (DF) generative model, which facilitates collaborative learning by combining input and output. Furthermore, we design a triplet model to capture correlations for real and AF attacked audios with real-generated and attacked-generated audios using auxiliary generative models. The proposed SHIELD strengthens the defense against generative AF attacks and achieves robust performance across various generative models. The proposed AF significantly reduces the average detection accuracy from 95.49% to 59.77% for ASVspoof2019, from 99.44% to 38.45% for In-the-Wild, and from 98.41% to 51.18% for HalfTruth for three different generative models. The proposed SHIELD mechanism is robust against AF attacks and achieves an average accuracy of 98.13%, 98.58%, and 99.57% in match, and 98.78%, 98.62%, and 98.85% in mismatch settings for the ASVspoof2019, In-the-Wild, and HalfTruth datasets, respectively.
☆ Orbis: Overcoming Challenges of Long-Horizon Prediction in Driving World Models
Existing world models for autonomous driving struggle with long-horizon generation and generalization to challenging scenarios. In this work, we develop a model using simple design choices, and without additional supervision or sensors, such as maps, depth, or multiple cameras. We show that our model yields state-of-the-art performance, despite having only 469M parameters and being trained on 280h of video data. It particularly stands out in difficult scenarios like turning maneuvers and urban traffic. We test whether discrete token models possibly have advantages over continuous models based on flow matching. To this end, we set up a hybrid tokenizer that is compatible with both approaches and allows for a side-by-side comparison. Our study concludes in favor of the continuous autoregressive model, which is less brittle on individual design choices and more powerful than the model built on discrete tokens. Code, models and qualitative results are publicly available at https://lmb-freiburg.github.io/orbis.github.io/.
comment: Project page: https://lmb-freiburg.github.io/orbis.github.io/
☆ Inverse Reinforcement Learning Meets Large Language Model Post-Training: Basics, Advances, and Opportunities
In the era of Large Language Models (LLMs), alignment has emerged as a fundamental yet challenging problem in the pursuit of more reliable, controllable, and capable machine intelligence. The recent success of reasoning models and conversational AI systems has underscored the critical role of reinforcement learning (RL) in enhancing these systems, driving increased research interest at the intersection of RL and LLM alignment. This paper provides a comprehensive review of recent advances in LLM alignment through the lens of inverse reinforcement learning (IRL), emphasizing the distinctions between RL techniques employed in LLM alignment and those in conventional RL tasks. In particular, we highlight the necessity of constructing neural reward models from human data and discuss the formal and practical implications of this paradigm shift. We begin by introducing fundamental concepts in RL to provide a foundation for readers unfamiliar with the field. We then examine recent advances in this research agenda, discussing key challenges and opportunities in conducting IRL for LLM alignment. Beyond methodological considerations, we explore practical aspects, including datasets, benchmarks, evaluation metrics, infrastructure, and computationally efficient training and inference techniques. Finally, we draw insights from the literature on sparse-reward RL to identify open questions and potential research directions. By synthesizing findings from diverse studies, we aim to provide a structured and critical overview of the field, highlight unresolved challenges, and outline promising future directions for improving LLM alignment through RL and IRL techniques.
☆ NonverbalTTS: A Public English Corpus of Text-Aligned Nonverbal Vocalizations with Emotion Annotations for Text-to-Speech
Current expressive speech synthesis models are constrained by the limited availability of open-source datasets containing diverse nonverbal vocalizations (NVs). In this work, we introduce NonverbalTTS (NVTTS), a 17-hour open-access dataset annotated with 10 types of NVs (e.g., laughter, coughs) and 8 emotional categories. The dataset is derived from popular sources, VoxCeleb and Expresso, using automated detection followed by human validation. We propose a comprehensive pipeline that integrates automatic speech recognition (ASR), NV tagging, emotion classification, and a fusion algorithm to merge transcriptions from multiple annotators. Fine-tuning open-source text-to-speech (TTS) models on the NVTTS dataset achieves parity with closed-source systems such as CosyVoice2, as measured by both human evaluation and automatic metrics, including speaker similarity and NV fidelity. By releasing NVTTS and its accompanying annotation guidelines, we address a key bottleneck in expressive TTS research. The dataset is available at https://huggingface.co/datasets/deepvk/NonverbalTTS.
☆ NGTM: Substructure-based Neural Graph Topic Model for Interpretable Graph Generation
Graph generation plays a pivotal role across numerous domains, including molecular design and knowledge graph construction. Although existing methods achieve considerable success in generating realistic graphs, their interpretability remains limited, often obscuring the rationale behind structural decisions. To address this challenge, we propose the Neural Graph Topic Model (NGTM), a novel generative framework inspired by topic modeling in natural language processing. NGTM represents graphs as mixtures of latent topics, each defining a distribution over semantically meaningful substructures, which facilitates explicit interpretability at both local and global scales. The generation process transparently integrates these topic distributions with a global structural variable, enabling clear semantic tracing of each generated graph. Experiments demonstrate that NGTM achieves competitive generation quality while uniquely enabling fine-grained control and interpretability, allowing users to tune structural features or induce biological properties through topic-level adjustments.
☆ Search for Z/2 eigenfunctions on the sphere using machine learning
We use machine learning to search for examples of Z/2 eigenfunctions on the 2-sphere. For this we created a multivalued version of a feedforward deep neural network, and we implemented it using the JAX library. We found Z/2 eigenfunctions for three cases: In the first two cases we fixed the branch points at the vertices of a tetrahedron and at a cube respectively. In a third case, we allowed the AI to move the branch points around and, in the end, it positioned the branch points at the vertices of a squashed tetrahedron.
comment: 14 pages, 12 pictures
☆ RS-TinyNet: Stage-wise Feature Fusion Network for Detecting Tiny Objects in Remote Sensing Images
Detecting tiny objects in remote sensing (RS) imagery has been a long-standing challenge due to their extremely limited spatial information, weak feature representations, and dense distributions across complex backgrounds. Despite numerous efforts devoted, mainstream detectors still underperform in such scenarios. To bridge this gap, we introduce RS-TinyNet, a multi-stage feature fusion and enhancement model explicitly tailored for RS tiny object detection in various RS scenarios. RS-TinyNet comes with two novel designs: tiny object saliency modeling and feature integrity reconstruction. Guided by these principles, we design three step-wise feature enhancement modules. Among them, the multi-dimensional collaborative attention (MDCA) module employs multi-dimensional attention to enhance the saliency of tiny objects. Additionally, the auxiliary reversible branch (ARB) and a progressive fusion detection head (PFDH) module are introduced to preserve information flow and fuse multi-level features to bridge semantic gaps and retain structural detail. Comprehensive experiments on public RS dataset AI-TOD show that our RS-TinyNet surpasses existing state-of-the-art (SOTA) detectors by 4.0% AP and 6.5% AP75. Evaluations on DIOR benchmark dataset further validate its superior detection performance in diverse RS scenarios. These results demonstrate that the proposed multi-stage feature fusion strategy offers an effective and practical solution for tiny object detection in complex RS environments.
☆ Deep Learning-Based Fetal Lung Segmentation from Diffusion-weighted MRI Images and Lung Maturity Evaluation for Fetal Growth Restriction
Fetal lung maturity is a critical indicator for predicting neonatal outcomes and the need for post-natal intervention, especially for pregnancies affected by fetal growth restriction. Intra-voxel incoherent motion analysis has shown promising results for non-invasive assessment of fetal lung development, but its reliance on manual segmentation is time-consuming, thus limiting its clinical applicability. In this work, we present an automated lung maturity evaluation pipeline for diffusion-weighted magnetic resonance images that consists of a deep learning-based fetal lung segmentation model and a model-fitting lung maturity assessment. A 3D nnU-Net model was trained on manually segmented images selected from the baseline frames of 4D diffusion-weighted MRI scans. The segmentation model demonstrated robust performance, yielding a mean Dice coefficient of 82.14%. Next, voxel-wise model fitting was performed based on both the nnU-Net-predicted and manual lung segmentations to quantify IVIM parameters reflecting tissue microstructure and perfusion. The results suggested no differences between the two. Our work shows that a fully automated pipeline is possible for supporting fetal lung maturity assessment and clinical decision-making.
☆ SemCSE: Semantic Contrastive Sentence Embeddings Using LLM-Generated Summaries For Scientific Abstracts
We introduce SemCSE, an unsupervised method for learning semantic embeddings of scientific texts. Building on recent advances in contrastive learning for text embeddings, our approach leverages LLM-generated summaries of scientific abstracts to train a model that positions semantically related summaries closer together in the embedding space. This resulting objective ensures that the model captures the true semantic content of a text, in contrast to traditional citation-based approaches that do not necessarily reflect semantic similarity. To validate this, we propose a novel benchmark designed to assess a model's ability to understand and encode the semantic content of scientific texts, demonstrating that our method enforces a stronger semantic separation within the embedding space. Additionally, we evaluate SemCSE on the comprehensive SciRepEval benchmark for scientific text embeddings, where it achieves state-of-the-art performance among models of its size, thus highlighting the benefits of a semantically focused training approach.
☆ Unsupervised Ground Metric Learning
Data classification without access to labeled samples remains a challenging problem. It usually depends on an appropriately chosen distance between features, a topic addressed in metric learning. Recently, Huizing, Cantini and Peyr\'e proposed to simultaneously learn optimal transport (OT) cost matrices between samples and features of the dataset. This leads to the task of finding positive eigenvectors of a certain nonlinear function that maps cost matrices to OT distances. Having this basic idea in mind, we consider both the algorithmic and the modeling part of unsupervised metric learning. First, we examine appropriate algorithms and their convergence. In particular, we propose to use the stochastic random function iteration algorithm and prove that it converges linearly for our setting, although our operators are not paracontractive as it was required for convergence so far. Second, we ask the natural question if the OT distance can be replaced by other distances. We show how Mahalanobis-like distances fit into our considerations. Further, we examine an approach via graph Laplacians. In contrast to the previous settings, we have just to deal with linear functions in the wanted matrices here, so that simple algorithms from linear algebra can be applied.
comment: 10 figures, 1 table
☆ Uncertainty-Aware Cross-Modal Knowledge Distillation with Prototype Learning for Multimodal Brain-Computer Interfaces
Electroencephalography (EEG) is a fundamental modality for cognitive state monitoring in brain-computer interfaces (BCIs). However, it is highly susceptible to intrinsic signal errors and human-induced labeling errors, which lead to label noise and ultimately degrade model performance. To enhance EEG learning, multimodal knowledge distillation (KD) has been explored to transfer knowledge from visual models with rich representations to EEG-based models. Nevertheless, KD faces two key challenges: modality gap and soft label misalignment. The former arises from the heterogeneous nature of EEG and visual feature spaces, while the latter stems from label inconsistencies that create discrepancies between ground truth labels and distillation targets. This paper addresses semantic uncertainty caused by ambiguous features and weakly defined labels. We propose a novel cross-modal knowledge distillation framework that mitigates both modality and label inconsistencies. It aligns feature semantics through a prototype-based similarity module and introduces a task-specific distillation head to resolve label-induced inconsistency in supervision. Experimental results demonstrate that our approach improves EEG-based emotion regression and classification performance, outperforming both unimodal and multimodal baselines on a public multimodal dataset. These findings highlight the potential of our framework for BCI applications.
☆ MUPAX: Multidimensional Problem Agnostic eXplainable AI
Robust XAI techniques should ideally be simultaneously deterministic, model agnostic, and guaranteed to converge. We propose MULTIDIMENSIONAL PROBLEM AGNOSTIC EXPLAINABLE AI (MUPAX), a deterministic, model agnostic explainability technique, with guaranteed convergency. MUPAX measure theoretic formulation gives principled feature importance attribution through structured perturbation analysis that discovers inherent input patterns and eliminates spurious relationships. We evaluate MUPAX on an extensive range of data modalities and tasks: audio classification (1D), image classification (2D), volumetric medical image analysis (3D), and anatomical landmark detection, demonstrating dimension agnostic effectiveness. The rigorous convergence guarantees extend to any loss function and arbitrary dimensions, making MUPAX applicable to virtually any problem context for AI. By contrast with other XAI methods that typically decrease performance when masking, MUPAX not only preserves but actually enhances model accuracy by capturing only the most important patterns of the original data. Extensive benchmarking against the state of the XAI art demonstrates MUPAX ability to generate precise, consistent and understandable explanations, a crucial step towards explainable and trustworthy AI systems. The source code will be released upon publication.
☆ DASViT: Differentiable Architecture Search for Vision Transformer IJCNN
Designing effective neural networks is a cornerstone of deep learning, and Neural Architecture Search (NAS) has emerged as a powerful tool for automating this process. Among the existing NAS approaches, Differentiable Architecture Search (DARTS) has gained prominence for its efficiency and ease of use, inspiring numerous advancements. Since the rise of Vision Transformers (ViT), researchers have applied NAS to explore ViT architectures, often focusing on macro-level search spaces and relying on discrete methods like evolutionary algorithms. While these methods ensure reliability, they face challenges in discovering innovative architectural designs, demand extensive computational resources, and are time-intensive. To address these limitations, we introduce Differentiable Architecture Search for Vision Transformer (DASViT), which bridges the gap in differentiable search for ViTs and uncovers novel designs. Experiments show that DASViT delivers architectures that break traditional Transformer encoder designs, outperform ViT-B/16 on multiple datasets, and achieve superior efficiency with fewer parameters and FLOPs.
comment: Accepted to the International Joint Conference on Neural Networks (IJCNN) 2025
☆ On statistical learning of graphs
We study PAC and online learnability of hypothesis classes formed by copies of a countably infinite graph G, where each copy is induced by permuting G's vertices. This corresponds to learning a graph's labeling, knowing its structure and label set. We consider classes where permutations move only finitely many vertices. Our main result shows that PAC learnability of all such finite-support copies implies online learnability of the full isomorphism type of G, and is equivalent to the condition of automorphic triviality. We also characterize graphs where copies induced by swapping two vertices are not learnable, using a relaxation of the extension property of the infinite random graph. Finally, we show that, for all G and k>2, learnability for k-vertex permutations is equivalent to that for 2-vertex permutations, yielding a four-class partition of infinite graphs, whose complexity we also determine using tools coming from both descriptive set theory and computability theory.
☆ The Power of Architecture: Deep Dive into Transformer Architectures for Long-Term Time Series Forecasting
Transformer-based models have recently become dominant in Long-term Time Series Forecasting (LTSF), yet the variations in their architecture, such as encoder-only, encoder-decoder, and decoder-only designs, raise a crucial question: What Transformer architecture works best for LTSF tasks? However, existing models are often tightly coupled with various time-series-specific designs, making it difficult to isolate the impact of the architecture itself. To address this, we propose a novel taxonomy that disentangles these designs, enabling clearer and more unified comparisons of Transformer architectures. Our taxonomy considers key aspects such as attention mechanisms, forecasting aggregations, forecasting paradigms, and normalization layers. Through extensive experiments, we uncover several key insights: bi-directional attention with joint-attention is most effective; more complete forecasting aggregation improves performance; and the direct-mapping paradigm outperforms autoregressive approaches. Furthermore, our combined model, utilizing optimal architectural choices, consistently outperforms several existing models, reinforcing the validity of our conclusions. We hope these findings offer valuable guidance for future research on Transformer architectural designs in LTSF. Our code is available at https://github.com/HALF111/TSF_architecture.
comment: 15 pages, 6 figures
☆ Confidence-Filtered Relevance (CFR): An Interpretable and Uncertainty-Aware Machine Learning Framework for Naturalness Assessment in Satellite Imagery
Protected natural areas play a vital role in ecological balance and ecosystem services. Monitoring these regions at scale using satellite imagery and machine learning is promising, but current methods often lack interpretability and uncertainty-awareness, and do not address how uncertainty affects naturalness assessment. In contrast, we propose Confidence-Filtered Relevance (CFR), a data-centric framework that combines LRP Attention Rollout with Deep Deterministic Uncertainty (DDU) estimation to analyze how model uncertainty influences the interpretability of relevance heatmaps. CFR partitions the dataset into subsets based on uncertainty thresholds, enabling systematic analysis of how uncertainty shapes the explanations of naturalness in satellite imagery. Applied to the AnthroProtect dataset, CFR assigned higher relevance to shrublands, forests, and wetlands, aligning with other research on naturalness assessment. Moreover, our analysis shows that as uncertainty increases, the interpretability of these relevance heatmaps declines and their entropy grows, indicating less selective and more ambiguous attributions. CFR provides a data-centric approach to assess the relevance of patterns to naturalness in satellite imagery based on their associated certainty.
☆ (Exhaustive) Symbolic Regression and model selection by minimum description length
Symbolic regression is the machine learning method for learning functions from data. After a brief overview of the symbolic regression landscape, I will describe the two main challenges that traditional algorithms face: they have an unknown (and likely significant) probability of failing to find any given good function, and they suffer from ambiguity and poorly-justified assumptions in their function-selection procedure. To address these I propose an exhaustive search and model selection by the minimum description length principle, which allows accuracy and complexity to be directly traded off by measuring each in units of information. I showcase the resulting publicly available Exhaustive Symbolic Regression algorithm on three open problems in astrophysics: the expansion history of the universe, the effective behaviour of gravity in galaxies and the potential of the inflaton field. In each case the algorithm identifies many functions superior to the literature standards. This general purpose methodology should find widespread utility in science and beyond.
comment: 15 pages, 4 figures; Invited review for the Royal Society Philosophical Transactions A special issue "Symbolic regression in the physical sciences"
☆ When Pattern-by-Pattern Works: Theoretical and Empirical Insights for Logistic Models with Missing Values
Predicting a response with partially missing inputs remains a challenging task even in parametric models, since parameter estimation in itself is not sufficient to predict on partially observed inputs. Several works study prediction in linear models. In this paper, we focus on logistic models, which present their own difficulties. From a theoretical perspective, we prove that a Pattern-by-Pattern strategy (PbP), which learns one logistic model per missingness pattern, accurately approximates Bayes probabilities in various missing data scenarios (MCAR, MAR and MNAR). Empirically, we thoroughly compare various methods (constant and iterative imputations, complete case analysis, PbP, and an EM algorithm) across classification, probability estimation, calibration, and parameter inference. Our analysis provides a comprehensive view on the logistic regression with missing values. It reveals that mean imputation can be used as baseline for low sample sizes, and improved performance is obtained via nonlinear multiple iterative imputation techniques with the labels (MICE.RF.Y). For large sample sizes, PbP is the best method for Gaussian mixtures, and we recommend MICE.RF.Y in presence of nonlinear features.
☆ Fault detection and diagnosis for the engine electrical system of a space launcher based on a temporal convolutional autoencoder and calibrated classifiers
In the context of the health monitoring for the next generation of reusable space launchers, we outline a first step toward developing an onboard fault detection and diagnostic capability for the electrical system that controls the engine valves. Unlike existing approaches in the literature, our solution is designed to meet a broader range of key requirements. This includes estimating confidence levels for predictions, detecting out-of-distribution (OOD) cases, and controlling false alarms. The proposed solution is based on a temporal convolutional autoencoder to automatically extract low-dimensional features from raw sensor data. Fault detection and diagnosis are respectively carried out using a binary and a multiclass classifier trained on the autoencoder latent and residual spaces. The classifiers are histogram-based gradient boosting models calibrated to output probabilities that can be interpreted as confidence levels. A relatively simple technique, based on inductive conformal anomaly detection, is used to identify OOD data. We leverage other simple yet effective techniques, such as cumulative sum control chart (CUSUM) to limit the false alarms, and threshold moving to address class imbalance in fault detection. The proposed framework is highly configurable and has been evaluated on simulated data, covering both nominal and anomalous operational scenarios. The results indicate that our solution is a promising first step, though testing with real data will be necessary to ensure that it achieves the required maturity level for operational use.
comment: 53 pages, 16 figures
☆ SMART: Relation-Aware Learning of Geometric Representations for Knowledge Graphs
Knowledge graph representation learning approaches provide a mapping between symbolic knowledge in the form of triples in a knowledge graph (KG) and their feature vectors. Knowledge graph embedding (KGE) models often represent relations in a KG as geometric transformations. Most state-of-the-art (SOTA) KGE models are derived from elementary geometric transformations (EGTs), such as translation, scaling, rotation, and reflection, or their combinations. These geometric transformations enable the models to effectively preserve specific structural and relational patterns of the KG. However, the current use of EGTs by KGEs remains insufficient without considering relation-specific transformations. Although recent models attempted to address this problem by ensembling SOTA baseline models in different ways, only a single or composite version of geometric transformations are used by such baselines to represent all the relations. In this paper, we propose a framework that evaluates how well each relation fits with different geometric transformations. Based on this ranking, the model can: (1) assign the best-matching transformation to each relation, or (2) use majority voting to choose one transformation type to apply across all relations. That is, the model learns a single relation-specific EGT in low dimensional vector space through an attention mechanism. Furthermore, we use the correlation between relations and EGTs, which are learned in a low dimension, for relation embeddings in a high dimensional vector space. The effectiveness of our models is demonstrated through comprehensive evaluations on three benchmark KGs as well as a real-world financial KG, witnessing a performance comparable to leading models
☆ Differential-informed Sample Selection Accelerates Multimodal Contrastive Learning
The remarkable success of contrastive-learning-based multimodal models has been greatly driven by training on ever-larger datasets with expensive compute consumption. Sample selection as an alternative efficient paradigm plays an important direction to accelerate the training process. However, recent advances on sample selection either mostly rely on an oracle model to offline select a high-quality coreset, which is limited in the cold-start scenarios, or focus on online selection based on real-time model predictions, which has not sufficiently or efficiently considered the noisy correspondence. To address this dilemma, we propose a novel Differential-Informed Sample Selection (DISSect) method, which accurately and efficiently discriminates the noisy correspondence for training acceleration. Specifically, we rethink the impact of noisy correspondence on contrastive learning and propose that the differential between the predicted correlation of the current model and that of a historical model is more informative to characterize sample quality. Based on this, we construct a robust differential-based sample selection and analyze its theoretical insights. Extensive experiments on three benchmark datasets and various downstream tasks demonstrate the consistent superiority of DISSect over current state-of-the-art methods. Source code is available at: https://github.com/MediaBrain-SJTU/DISSect.
☆ Teach Old SAEs New Domain Tricks with Boosting
Sparse Autoencoders have emerged as powerful tools for interpreting the internal representations of Large Language Models, yet they often fail to capture domain-specific features not prevalent in their training corpora. This paper introduces a residual learning approach that addresses this feature blindness without requiring complete retraining. We propose training a secondary SAE specifically to model the reconstruction error of a pretrained SAE on domain-specific texts, effectively capturing features missed by the primary model. By summing the outputs of both models during inference, we demonstrate significant improvements in both LLM cross-entropy and explained variance metrics across multiple specialized domains. Our experiments show that this method efficiently incorporates new domain knowledge into existing SAEs while maintaining their performance on general tasks. This approach enables researchers to selectively enhance SAE interpretability for specific domains of interest, opening new possibilities for targeted mechanistic interpretability of LLMs.
☆ Variance-Based Pruning for Accelerating and Compressing Trained Networks ICCV
Increasingly expensive training of ever larger models such as Vision Transfomers motivate reusing the vast library of already trained state-of-the-art networks. However, their latency, high computational costs and memory demands pose significant challenges for deployment, especially on resource-constrained hardware. While structured pruning methods can reduce these factors, they often require costly retraining, sometimes for up to hundreds of epochs, or even training from scratch to recover the lost accuracy resulting from the structural modifications. Maintaining the provided performance of trained models after structured pruning and thereby avoiding extensive retraining remains a challenge. To solve this, we introduce Variance-Based Pruning, a simple and structured one-shot pruning technique for efficiently compressing networks, with minimal finetuning. Our approach first gathers activation statistics, which are used to select neurons for pruning. Simultaneously the mean activations are integrated back into the model to preserve a high degree of performance. On ImageNet-1k recognition tasks, we demonstrate that directly after pruning DeiT-Base retains over 70% of its original performance and requires only 10 epochs of fine-tuning to regain 99% of the original accuracy while simultaneously reducing MACs by 35% and model size by 36%, thus speeding up the model by 1.44x.
comment: Accepted at IEEE/CVF International Conference on Computer Vision (ICCV) 2025
☆ FedGA: A Fair Federated Learning Framework Based on the Gini Coefficient
Fairness has emerged as one of the key challenges in federated learning. In horizontal federated settings, data heterogeneity often leads to substantial performance disparities across clients, raising concerns about equitable model behavior. To address this issue, we propose FedGA, a fairness-aware federated learning algorithm. We first employ the Gini coefficient to measure the performance disparity among clients. Based on this, we establish a relationship between the Gini coefficient $G$ and the update scale of the global model ${U_s}$, and use this relationship to adaptively determine the timing of fairness intervention. Subsequently, we dynamically adjust the aggregation weights according to the system's real-time fairness status, enabling the global model to better incorporate information from clients with relatively poor performance.We conduct extensive experiments on the Office-Caltech-10, CIFAR-10, and Synthetic datasets. The results show that FedGA effectively improves fairness metrics such as variance and the Gini coefficient, while maintaining strong overall performance, demonstrating the effectiveness of our approach.
☆ A Distributed Generative AI Approach for Heterogeneous Multi-Domain Environments under Data Sharing constraints
Federated Learning has gained increasing attention for its ability to enable multiple nodes to collaboratively train machine learning models without sharing their raw data. At the same time, Generative AI -- particularly Generative Adversarial Networks (GANs) -- have achieved remarkable success across a wide range of domains, such as healthcare, security, and Image Generation. However, training generative models typically requires large datasets and significant computational resources, which are often unavailable in real-world settings. Acquiring such resources can be costly and inefficient, especially when many underutilized devices -- such as IoT devices and edge devices -- with varying capabilities remain idle. Moreover, obtaining large datasets is challenging due to privacy concerns and copyright restrictions, as most devices are unwilling to share their data. To address these challenges, we propose a novel approach for decentralized GAN training that enables the utilization of distributed data and underutilized, low-capability devices while not sharing data in its raw form. Our approach is designed to tackle key challenges in decentralized environments, combining KLD-weighted Clustered Federated Learning to address the issues of data heterogeneity and multi-domain datasets, with Heterogeneous U-Shaped split learning to tackle the challenge of device heterogeneity under strict data sharing constraints -- ensuring that no labels or raw data, whether real or synthetic, are ever shared between nodes. Experimental results shows that our approach demonstrates consistent and significant improvements across key performance metrics, where it achieves 1.1x -- 2.2x higher image generation scores, an average 10% boost in classification metrics (up to 50% in multi-domain non-IID settings), in much lower latency compared to several benchmarks. Find our code at https://github.com/youssefga28/HuSCF-GAN.
☆ WaveletInception Networks for Drive-by Vibration-Based Infrastructure Health Monitoring
This paper presents a novel deep learning-based framework for infrastructure health monitoring using drive-by vibration response signals. Recognizing the importance of spectral and temporal information, we introduce the WaveletInception-BiLSTM network. The WaveletInception feature extractor utilizes a Learnable Wavelet Packet Transform (LWPT) as the stem for extracting vibration signal features, incorporating spectral information in the early network layers. This is followed by 1D Inception networks that extract multi-scale, high-level features at deeper layers. The extracted vibration signal features are then integrated with operational conditions via a Long Short-term Memory (LSTM) layer. The resulting feature extraction network effectively analyzes drive-by vibration signals across various measurement speeds without preprocessing and uses LSTM to capture interrelated temporal dependencies among different modes of information and to create feature vectors for health condition estimation. The estimator head is designed with a sequential modeling architecture using bidirectional LSTM (BiLSTM) networks, capturing bi-directional temporal relationships from drive-by measurements. This architecture allows for a high-resolution, beam-level assessment of infrastructure health conditions. A case study focusing on railway track stiffness estimation with simulated drive-by vibration signals shows that the model significantly outperforms state-of-the-art methods in estimating railway ballast and railpad stiffness parameters. Results underscore the potential of this approach for accurate, localized, and fully automated drive-by infrastructure health monitoring.
☆ Investigating Forecasting Models for Pandemic Infections Using Heterogeneous Data Sources: A 2-year Study with COVID-19
Emerging in December 2019, the COVID-19 pandemic caused widespread health, economic, and social disruptions. Rapid global transmission overwhelmed healthcare systems, resulting in high infection rates, hospitalisations, and fatalities. To minimise the spread, governments implemented several non-pharmaceutical interventions like lockdowns and travel restrictions. While effective in controlling transmission, these measures also posed significant economic and societal challenges. Although the WHO declared COVID-19 no longer a global health emergency in May 2023, its impact persists, shaping public health strategies. The vast amount of data collected during the pandemic offers valuable insights into disease dynamics, transmission, and intervention effectiveness. Leveraging these insights can improve forecasting models, enhancing preparedness and response to future outbreaks while mitigating their social and economic impact. This paper presents a large-scale case study on COVID-19 forecasting in Cyprus, utilising a two-year dataset that integrates epidemiological data, vaccination records, policy measures, and weather conditions. We analyse infection trends, assess forecasting performance, and examine the influence of external factors on disease dynamics. The insights gained contribute to improved pandemic preparedness and response strategies.
comment: Keywords: epidemiology, pandemic forecasting, COVID-19, infections, machine learning Accepted: IEEE Conference on Computational Intelligence in Bioinformatics and Computational Biology (CIBCB) 2025
☆ Demographic-aware fine-grained classification of pediatric wrist fractures
Wrist pathologies are frequently observed, particularly among children who constitute the majority of fracture cases. However, diagnosing these conditions is time-consuming and requires specialized expertise. Computer vision presents a promising avenue, contingent upon the availability of extensive datasets, a notable challenge in medical imaging. Therefore, reliance solely on one modality, such as images, proves inadequate, especially in an era of diverse and plentiful data types. In this study, we employ a multifaceted approach to address the challenge of recognizing wrist pathologies using an extremely limited dataset. Initially, we approach the problem as a fine-grained recognition task, aiming to identify subtle X-ray pathologies that conventional CNNs overlook. Secondly, we enhance network performance by fusing patient metadata with X-ray images. Thirdly, rather than pre-training on a coarse-grained dataset like ImageNet, we utilize weights trained on a fine-grained dataset. While metadata integration has been used in other medical domains, this is a novel application for wrist pathologies. Our results show that a fine-grained strategy and metadata integration improve diagnostic accuracy by 2% with a limited dataset and by over 10% with a larger fracture-focused dataset.
☆ A Spectral Interpretation of Redundancy in a Graph Reservoir ICANN 2025
Reservoir computing has been successfully applied to graphs as a preprocessing method to improve the training efficiency of Graph Neural Networks (GNNs). However, a common issue that arises when repeatedly applying layer operators on graphs is over-smoothing, which consists in the convergence of graph signals toward low-frequency components of the graph Laplacian. This work revisits the definition of the reservoir in the Multiresolution Reservoir Graph Neural Network (MRGNN), a spectral reservoir model, and proposes a variant based on a Fairing algorithm originally introduced in the field of surface design in computer graphics. This algorithm provides a pass-band spectral filter that allows smoothing without shrinkage, and it can be adapted to the graph setting through the Laplacian operator. Given its spectral formulation, this method naturally connects to GNN architectures for tasks where smoothing, when properly controlled, can be beneficial,such as graph classification. The core contribution of the paper lies in the theoretical analysis of the algorithm from a random walks perspective. In particular, it shows how tuning the spectral coefficients can be interpreted as modulating the contribution of redundant random walks. Exploratory experiments based on the MRGNN architecture illustrate the potential of this approach and suggest promising directions for future research.
comment: This paper has been accepted for presentation at the 3rd International Workshop on Reservoir Computing (RC 2025) at ICANN 2025
☆ cIDIR: Conditioned Implicit Neural Representation for Regularized Deformable Image Registration
Regularization is essential in deformable image registration (DIR) to ensure that the estimated Deformation Vector Field (DVF) remains smooth, physically plausible, and anatomically consistent. However, fine-tuning regularization parameters in learning-based DIR frameworks is computationally expensive, often requiring multiple training iterations. To address this, we propose cIDI, a novel DIR framework based on Implicit Neural Representations (INRs) that conditions the registration process on regularization hyperparameters. Unlike conventional methods that require retraining for each regularization hyperparameter setting, cIDIR is trained over a prior distribution of these hyperparameters, then optimized over the regularization hyperparameters by using the segmentations masks as an observation. Additionally, cIDIR models a continuous and differentiable DVF, enabling seamless integration of advanced regularization techniques via automatic differentiation. Evaluated on the DIR-LAB dataset, $\operatorname{cIDIR}$ achieves high accuracy and robustness across the dataset.
☆ Insights into a radiology-specialised multimodal large language model with sparse autoencoders ICML 2025
Interpretability can improve the safety, transparency and trust of AI models, which is especially important in healthcare applications where decisions often carry significant consequences. Mechanistic interpretability, particularly through the use of sparse autoencoders (SAEs), offers a promising approach for uncovering human-interpretable features within large transformer-based models. In this study, we apply Matryoshka-SAE to the radiology-specialised multimodal large language model, MAIRA-2, to interpret its internal representations. Using large-scale automated interpretability of the SAE features, we identify a range of clinically relevant concepts - including medical devices (e.g., line and tube placements, pacemaker presence), pathologies such as pleural effusion and cardiomegaly, longitudinal changes and textual features. We further examine the influence of these features on model behaviour through steering, demonstrating directional control over generations with mixed success. Our results reveal practical and methodological challenges, yet they offer initial insights into the internal concepts learned by MAIRA-2 - marking a step toward deeper mechanistic understanding and interpretability of a radiology-adapted multimodal large language model, and paving the way for improved model transparency. We release the trained SAEs and interpretations: https://huggingface.co/microsoft/maira-2-sae.
comment: Actionable Interpretability Workshop at ICML 2025. 24 pages, 7 figures, 5 tables
☆ Probabilistic Soundness Guarantees in LLM Reasoning Chains
In reasoning chains generated by large language models (LLMs), initial errors often propagate and undermine the reliability of the final conclusion. Current LLM-based error detection methods often fail to detect propagated errors because they do not properly account for how earlier errors might corrupt judgments of downstream reasoning. To better detect such propagated errors, we introduce Autoregressive Reasoning Entailment Stability (ARES), a novel probabilistic framework that prevents error propagation by judging each claim based only on previously-assessed sound premises. This inductive method yields a nuanced score for each step and provides certified statistical guarantees of its soundness, rather than a brittle binary label. ARES achieves state-of-the-art performance across four benchmarks (72.1% Macro-F1, +8.2 points) and demonstrates superior robustness on very long synthetic reasoning chains, where it excels at detecting propagated errors (90.3% F1, +27.6 points).
☆ MC$^2$A: Enabling Algorithm-Hardware Co-Design for Efficient Markov Chain Monte Carlo Acceleration
An increasing number of applications are exploiting sampling-based algorithms for planning, optimization, and inference. The Markov Chain Monte Carlo (MCMC) algorithms form the computational backbone of this emerging branch of machine learning. Unfortunately, the high computational cost limits their feasibility for large-scale problems and real-world applications, and the existing MCMC acceleration solutions are either limited in hardware flexibility or fail to maintain efficiency at the system level across a variety of end-to-end applications. This paper introduces \textbf{MC$^2$A}, an algorithm-hardware co-design framework, enabling efficient and flexible optimization for MCMC acceleration. Firstly, \textbf{MC$^2$A} analyzes the MCMC workload diversity through an extension of the processor performance roofline model with a 3rd dimension to derive the optimal balance between the compute, sampling and memory parameters. Secondly, \textbf{MC$^2$A} proposes a parametrized hardware accelerator architecture with flexible and efficient support of MCMC kernels with a pipeline of ISA-programmable tree-structured processing units, reconfigurable samplers and a crossbar interconnect to support irregular access. Thirdly, the core of \textbf{MC$^2$A} is powered by a novel Gumbel sampler that eliminates exponential and normalization operations. In the end-to-end case study, \textbf{MC$^2$A} achieves an overall {$307.6\times$, $1.4\times$, $2.0\times$, $84.2\times$} speedup compared to the CPU, GPU, TPU and state-of-the-art MCMC accelerator. Evaluated on various representative MCMC workloads, this work demonstrates and exploits the feasibility of general hardware acceleration to popularize MCMC-based solutions in diverse application domains.
comment: 14 pages, 15 figures, IEEE journal paper
☆ DMQ: Dissecting Outliers of Diffusion Models for Post-Training Quantization ICCV 2025
Diffusion models have achieved remarkable success in image generation but come with significant computational costs, posing challenges for deployment in resource-constrained environments. Recent post-training quantization (PTQ) methods have attempted to mitigate this issue by focusing on the iterative nature of diffusion models. However, these approaches often overlook outliers, leading to degraded performance at low bit-widths. In this paper, we propose a DMQ which combines Learned Equivalent Scaling (LES) and channel-wise Power-of-Two Scaling (PTS) to effectively address these challenges. Learned Equivalent Scaling optimizes channel-wise scaling factors to redistribute quantization difficulty between weights and activations, reducing overall quantization error. Recognizing that early denoising steps, despite having small quantization errors, crucially impact the final output due to error accumulation, we incorporate an adaptive timestep weighting scheme to prioritize these critical steps during learning. Furthermore, identifying that layers such as skip connections exhibit high inter-channel variance, we introduce channel-wise Power-of-Two Scaling for activations. To ensure robust selection of PTS factors even with small calibration set, we introduce a voting algorithm that enhances reliability. Extensive experiments demonstrate that our method significantly outperforms existing works, especially at low bit-widths such as W4A6 (4-bit weight, 6-bit activation) and W4A8, maintaining high image generation quality and model stability. The code is available at https://github.com/LeeDongYeun/dmq.
comment: Accepted by ICCV 2025
☆ From a Mixed-Policy Perspective: Improving Differentiable Automatic Post-editing Optimization
This paper introduces two novel modifications to the Differentiable Automatic Post-editing Optimization (DAPO) algorithm, approached from a mixed-policy perspective. Standard policy gradient methods can suffer from instability and sample inefficiency, particularly in sparse reward settings. To address this, we first propose a method that incorporates a pre-trained, stable guiding policy ($\piphi$) to provide off-policy experience, thereby regularizing the training of the target policy ($\pion$). This approach improves training stability and convergence speed by adaptively adjusting the learning step size. Secondly, we extend this idea to re-utilize zero-reward samples, which are often discarded by dynamic sampling strategies like DAPO's. By treating these samples as a distinct batch guided by the expert policy, we further enhance sample efficiency. We provide a theoretical analysis for both methods, demonstrating that their objective functions converge to the optimal solution within the established theoretical framework of reinforcement learning. The proposed mixed-policy framework effectively balances exploration and exploitation, promising more stable and efficient policy optimization.
☆ Trace Reconstruction with Language Models
The general trace reconstruction problem seeks to recover an original sequence from its noisy copies independently corrupted by deletions, insertions, and substitutions. This problem arises in applications such as DNA data storage, a promising storage medium due to its high information density and longevity. However, errors introduced during DNA synthesis, storage, and sequencing require correction through algorithms and codes, with trace reconstruction often used as part of the data retrieval process. In this work, we propose TReconLM, which leverages language models trained on next-token prediction for trace reconstruction. We pretrain language models on synthetic data and fine-tune on real-world data to adapt to technology-specific error patterns. TReconLM outperforms state-of-the-art trace reconstruction algorithms, including prior deep learning approaches, recovering a substantially higher fraction of sequences without error.
☆ Robust Explanations Through Uncertainty Decomposition: A Path to Trustworthier AI
Recent advancements in machine learning have emphasized the need for transparency in model predictions, particularly as interpretability diminishes when using increasingly complex architectures. In this paper, we propose leveraging prediction uncertainty as a complementary approach to classical explainability methods. Specifically, we distinguish between aleatoric (data-related) and epistemic (model-related) uncertainty to guide the selection of appropriate explanations. Epistemic uncertainty serves as a rejection criterion for unreliable explanations and, in itself, provides insight into insufficient training (a new form of explanation). Aleatoric uncertainty informs the choice between feature-importance explanations and counterfactual explanations. This leverages a framework of explainability methods driven by uncertainty quantification and disentanglement. Our experiments demonstrate the impact of this uncertainty-aware approach on the robustness and attainability of explanations in both traditional machine learning and deep learning scenarios.
☆ LaViPlan : Language-Guided Visual Path Planning with RLVR
Out-of-distribution (OOD) scenarios in autonomous driving refer to situations that deviate from the training domain, often leading to unexpected and potentially hazardous behavior from planners that lack prior exposure to such cases. Recently, Vision-Language Models (VLMs) have been introduced into autonomous driving research for their promising generalization capabilities in OOD settings. Early studies demonstrated that VLMs could recognize OOD scenarios and generate user-level decisions such as "go straight" or "turn right." However, a new challenge has emerged due to the misalignment between the VLM's high-level decisions or visual reasoning expressed in language, and the low-level predicted trajectories interpreted as actions. In this paper, we propose LaViPlan, a framework that leverages Reinforcement Learning with Verifiable Rewards (RLVR) to optimize VLMs using planning-oriented metrics. This approach addresses the vision-language-action misalignment observed in existing VLMs fine-tuned via supervised learning, which can recognize driving scenarios but often produce context-unaware decisions. Experimental results demonstrate that our method improves situational awareness and decision-making under OOD conditions, highlighting its potential to mitigate the misalignment issue. This work introduces a promising post-training paradigm for VLM agents in the context of autonomous driving.
comment: 11 pages, 6 figures
☆ Fremer: Lightweight and Effective Frequency Transformer for Workload Forecasting in Cloud Services
Workload forecasting is pivotal in cloud service applications, such as auto-scaling and scheduling, with profound implications for operational efficiency. Although Transformer-based forecasting models have demonstrated remarkable success in general tasks, their computational efficiency often falls short of the stringent requirements in large-scale cloud environments. Given that most workload series exhibit complicated periodic patterns, addressing these challenges in the frequency domain offers substantial advantages. To this end, we propose Fremer, an efficient and effective deep forecasting model. Fremer fulfills three critical requirements: it demonstrates superior efficiency, outperforming most Transformer-based forecasting models; it achieves exceptional accuracy, surpassing all state-of-the-art (SOTA) models in workload forecasting; and it exhibits robust performance for multi-period series. Furthermore, we collect and open-source four high-quality, open-source workload datasets derived from ByteDance's cloud services, encompassing workload data from thousands of computing instances. Extensive experiments on both our proprietary datasets and public benchmarks demonstrate that Fremer consistently outperforms baseline models, achieving average improvements of 5.5% in MSE, 4.7% in MAE, and 8.6% in SMAPE over SOTA models, while simultaneously reducing parameter scale and computational costs. Additionally, in a proactive auto-scaling test based on Kubernetes, Fremer improves average latency by 18.78% and reduces resource consumption by 2.35%, underscoring its practical efficacy in real-world applications.
comment: 12 pages, 11 figures
☆ Learning to Reject Low-Quality Explanations via User Feedback
Machine Learning predictors are increasingly being employed in high-stakes applications such as credit scoring. Explanations help users unpack the reasons behind their predictions, but are not always "high quality''. That is, end-users may have difficulty interpreting or believing them, which can complicate trust assessment and downstream decision-making. We argue that classifiers should have the option to refuse handling inputs whose predictions cannot be explained properly and introduce a framework for learning to reject low-quality explanations (LtX) in which predictors are equipped with a rejector that evaluates the quality of explanations. In this problem setting, the key challenges are how to properly define and assess explanation quality and how to design a suitable rejector. Focusing on popular attribution techniques, we introduce ULER (User-centric Low-quality Explanation Rejector), which learns a simple rejector from human ratings and per-feature relevance judgments to mirror human judgments of explanation quality. Our experiments show that ULER outperforms both state-of-the-art and explanation-aware learning to reject strategies at LtX on eight classification and regression benchmarks and on a new human-annotated dataset, which we will publicly release to support future research.
☆ Generalist Bimanual Manipulation via Foundation Video Diffusion Models
Bimanual robotic manipulation, which involves the coordinated control of two robotic arms, is foundational for solving challenging tasks. Despite recent progress in general-purpose manipulation, data scarcity and embodiment heterogeneity remain serious obstacles to further scaling up in bimanual settings. In this paper, we introduce VIdeo Diffusion for Action Reasoning (VIDAR), a two-stage framework that leverages large-scale, diffusion-based video pre-training and a novel masked inverse dynamics model for action prediction. We pre-train the video diffusion model on 750K multi-view videos from three real-world bimanual robot platforms, utilizing a unified observation space that encodes robot, camera, task, and scene contexts. Our masked inverse dynamics model learns masks to extract action-relevant information from generated trajectories without requiring pixel-level labels, and the masks can effectively generalize to unseen backgrounds. Our experiments demonstrate that with only 20 minutes of human demonstrations on an unseen robot platform (only 1% of typical data requirements), VIDAR generalizes to unseen tasks and backgrounds with strong semantic understanding, surpassing state-of-the-art methods. Our findings highlight the potential of video foundation models, coupled with masked action prediction, to enable scalable and generalizable robotic manipulation in diverse real-world settings.
☆ VAR-MATH: Probing True Mathematical Reasoning in Large Language Models via Symbolic Multi-Instance Benchmarks
Recent advances in reinforcement learning (RL) have led to substantial improvements in the mathematical reasoning abilities of large language models (LLMs), as measured by standard benchmarks. However, these gains often persist even when models are trained with flawed signals, such as random or inverted rewards, raising a fundamental question: do such improvements reflect true reasoning, or are they merely artifacts of overfitting to benchmark-specific patterns? To address this question, we take an evaluation-centric perspective and identify two critical shortcomings in existing protocols. First, \emph{benchmark contamination} arises from the public availability of test problems, increasing the risk of data leakage. Second, \emph{evaluation fragility} stems from the reliance on single-instance assessments, which are highly sensitive to stochastic outputs and fail to capture reasoning consistency. To overcome these limitations, we introduce {VAR-MATH}, a symbolic evaluation framework designed to probe genuine reasoning ability. By converting fixed numerical problems into symbolic templates and requiring models to solve multiple instantiations of each, VAR-MATH enforces consistent reasoning across structurally equivalent variants, thereby mitigating contamination and improving evaluation robustness. We apply VAR-MATH to transform two popular benchmarks, AMC23 and AIME24, into their symbolic counterparts, VAR-AMC23 and VAR-AIME24. Experimental results reveal substantial performance drops for RL-trained models on the variabilized versions, especially for smaller models, with average declines of 48.0\% on AMC23 and 58.3\% on AIME24. These findings suggest that many existing RL methods rely on superficial heuristics and fail to generalize beyond specific numerical forms. Overall, VAR-MATH offers a principled, contamination-resistant evaluation paradigm for mathematical reasoning.
☆ Autonomous Resource Management in Microservice Systems via Reinforcement Learning
This paper proposes a reinforcement learning-based method for microservice resource scheduling and optimization, aiming to address issues such as uneven resource allocation, high latency, and insufficient throughput in traditional microservice architectures. In microservice systems, as the number of services and the load increase, efficiently scheduling and allocating resources such as computing power, memory, and storage becomes a critical research challenge. To address this, the paper employs an intelligent scheduling algorithm based on reinforcement learning. Through the interaction between the agent and the environment, the resource allocation strategy is continuously optimized. In the experiments, the paper considers different resource conditions and load scenarios, evaluating the proposed method across multiple dimensions, including response time, throughput, resource utilization, and cost efficiency. The experimental results show that the reinforcement learning-based scheduling method significantly improves system response speed and throughput under low load and high concurrency conditions, while also optimizing resource utilization and reducing energy consumption. Under multi-dimensional resource conditions, the proposed method can consider multiple objectives and achieve optimized resource scheduling. Compared to traditional static resource allocation methods, the reinforcement learning model demonstrates stronger adaptability and optimization capability. It can adjust resource allocation strategies in real time, thereby maintaining good system performance in dynamically changing load and resource environments.
☆ Bayesian Modeling and Estimation of Linear Time-Variant Systems using Neural Networks and Gaussian Processes
The identification of Linear Time-Variant (LTV) systems from input-output data is a fundamental yet challenging ill-posed inverse problem. This work introduces a unified Bayesian framework that models the system's impulse response, $h(t, \tau)$, as a stochastic process. We decompose the response into a posterior mean and a random fluctuation term, a formulation that provides a principled approach for quantifying uncertainty and naturally defines a new, useful system class we term Linear Time-Invariant in Expectation (LTIE). To perform inference, we leverage modern machine learning techniques, including Bayesian neural networks and Gaussian Processes, using scalable variational inference. We demonstrate through a series of experiments that our framework can robustly infer the properties of an LTI system from a single noisy observation, show superior data efficiency compared to classical methods in a simulated ambient noise tomography problem, and successfully track a continuously varying LTV impulse response by using a structured Gaussian Process prior. This work provides a flexible and robust methodology for uncertainty-aware system identification in dynamic environments.
☆ Topology-Aware Activation Functions in Neural Networks
This study explores novel activation functions that enhance the ability of neural networks to manipulate data topology during training. Building on the limitations of traditional activation functions like $\mathrm{ReLU}$, we propose $\mathrm{SmoothSplit}$ and $\mathrm{ParametricSplit}$, which introduce topology "cutting" capabilities. These functions enable networks to transform complex data manifolds effectively, improving performance in scenarios with low-dimensional layers. Through experiments on synthetic and real-world datasets, we demonstrate that $\mathrm{ParametricSplit}$ outperforms traditional activations in low-dimensional settings while maintaining competitive performance in higher-dimensional ones. Our findings highlight the potential of topology-aware activation functions in advancing neural network architectures. The code is available via https://github.com/Snopoff/Topology-Aware-Activations.
comment: Accepted to ESANN 2025. Published in the ESANN 2025 proceedings
☆ An Investigation of Ear-EEG Signals for a Novel Biometric Authentication System
This work explores the feasibility of biometric authentication using EEG signals acquired through in-ear devices, commonly referred to as ear-EEG. Traditional EEG-based biometric systems, while secure, often suffer from low usability due to cumbersome scalp-based electrode setups. In this study, we propose a novel and practical framework leveraging ear-EEG signals as a user-friendly alternative for everyday biometric authentication. The system extracts an original combination of temporal and spectral features from ear-EEG signals and feeds them into a fully connected deep neural network for subject identification. Experimental results on the only currently available ear-EEG dataset suitable for different purposes, including biometric authentication, demonstrate promising performance, with an average accuracy of 82\% in a subject identification scenario. These findings confirm the potential of ear-EEG as a viable and deployable direction for next-generation real-world biometric systems.
☆ WhoFi: Deep Person Re-Identification via Wi-Fi Channel Signal Encoding
Person Re-Identification is a key and challenging task in video surveillance. While traditional methods rely on visual data, issues like poor lighting, occlusion, and suboptimal angles often hinder performance. To address these challenges, we introduce WhoFi, a novel pipeline that utilizes Wi-Fi signals for person re-identification. Biometric features are extracted from Channel State Information (CSI) and processed through a modular Deep Neural Network (DNN) featuring a Transformer-based encoder. The network is trained using an in-batch negative loss function to learn robust and generalizable biometric signatures. Experiments on the NTU-Fi dataset show that our approach achieves competitive results compared to state-of-the-art methods, confirming its effectiveness in identifying individuals via Wi-Fi signals.
☆ Supervised Fine Tuning on Curated Data is Reinforcement Learning (and can be improved)
Behavior Cloning (BC) on curated (or filtered) data is the predominant paradigm for supervised fine-tuning (SFT) of large language models; as well as for imitation learning of control policies. Here, we draw on a connection between this successful strategy and the theory and practice of finding optimal policies via Reinforcement Learning (RL). Building on existing literature, we clarify that SFT can be understood as maximizing a lower bound on the RL objective in a sparse reward setting. Giving support to its often observed good performance. From this viewpoint, we realize that a small modification to SFT leads to an importance weighted variant that behaves closer to training with RL as it: i) optimizes a tighter bound to the RL objective and, ii) can improve performance compared to SFT on curated data. We refer to this variant as importance weighted supervised fine-tuning (iw-SFT). We show that it is easy to implement and can be further generalized to training with quality scored data. The resulting SFT variants are competitive with more advanced RL algorithms for large language models and for training policies in continuous control tasks. For example achieving 66.7% on the AIME 2024 dataset.
comment: See project website for details and code at: https://independentresearch.ai/posts/iwsft
Transformer-Based Person Identification via Wi-Fi CSI Amplitude and Phase Perturbations
Wi-Fi sensing is gaining momentum as a non-intrusive and privacy-preserving alternative to vision-based systems for human identification. However, person identification through wireless signals, particularly without user motion, remains largely unexplored. Most prior wireless-based approaches rely on movement patterns, such as walking gait, to extract biometric cues. In contrast, we propose a transformer-based method that identifies individuals from Channel State Information (CSI) recorded while the subject remains stationary. CSI captures fine-grained amplitude and phase distortions induced by the unique interaction between the human body and the radio signal. To support evaluation, we introduce a dataset acquired with ESP32 devices in a controlled indoor environment, featuring six participants observed across multiple orientations. A tailored preprocessing pipeline, including outlier removal, smoothing, and phase calibration, enhances signal quality. Our dual-branch transformer architecture processes amplitude and phase modalities separately and achieves 99.82\% classification accuracy, outperforming convolutional and multilayer perceptron baselines. These results demonstrate the discriminative potential of CSI perturbations, highlighting their capacity to encode biometric traits in a consistent manner. They further confirm the viability of passive, device-free person identification using low-cost commodity Wi-Fi hardware in real-world settings.
☆ A Kernel Distribution Closeness Testing
The distribution closeness testing (DCT) assesses whether the distance between a distribution pair is at least $\epsilon$-far. Existing DCT methods mainly measure discrepancies between a distribution pair defined on discrete one-dimensional spaces (e.g., using total variation), which limits their applications to complex data (e.g., images). To extend DCT to more types of data, a natural idea is to introduce maximum mean discrepancy (MMD), a powerful measurement of the distributional discrepancy between two complex distributions, into DCT scenarios. However, we find that MMD's value can be the same for many pairs of distributions that have different norms in the same reproducing kernel Hilbert space (RKHS), making MMD less informative when assessing the closeness levels for multiple distribution pairs. To mitigate the issue, we design a new measurement of distributional discrepancy, norm-adaptive MMD (NAMMD), which scales MMD's value using the RKHS norms of distributions. Based on the asymptotic distribution of NAMMD, we finally propose the NAMMD-based DCT to assess the closeness levels of a distribution pair. Theoretically, we prove that NAMMD-based DCT has higher test power compared to MMD-based DCT, with bounded type-I error, which is also validated by extensive experiments on many types of data (e.g., synthetic noise, real images). Furthermore, we also apply the proposed NAMMD for addressing the two-sample testing problem and find NAMMD-based two-sample test has higher test power than the MMD-based two-sample test in both theory and experiments.
☆ Bridging the Gap: Leveraging Retrieval-Augmented Generation to Better Understand Public Concerns about Vaccines
Vaccine hesitancy threatens public health, leading to delayed or rejected vaccines. Social media is a vital source for understanding public concerns, and traditional methods like topic modelling often struggle to capture nuanced opinions. Though trained for query answering, large Language Models (LLMs) often miss current events and community concerns. Additionally, hallucinations in LLMs can compromise public health communication. To address these limitations, we developed a tool (VaxPulse Query Corner) using the Retrieval Augmented Generation technique. It addresses complex queries about public vaccine concerns on various online platforms, aiding public health administrators and stakeholders in understanding public concerns and implementing targeted interventions to boost vaccine confidence. Analysing 35,103 Shingrix social media posts, it achieved answer faithfulness (0.96) and relevance (0.94).
☆ Understanding the Evolution of the Neural Tangent Kernel at the Edge of Stability
The study of Neural Tangent Kernels (NTKs) in deep learning has drawn increasing attention in recent years. NTKs typically actively change during training and are related to feature learning. In parallel, recent work on Gradient Descent (GD) has found a phenomenon called Edge of Stability (EoS), in which the largest eigenvalue of the NTK oscillates around a value inversely proportional to the step size. However, although follow-up works have explored the underlying mechanism of such eigenvalue behavior in depth, the understanding of the behavior of the NTK eigenvectors during EoS is still missing. This paper examines the dynamics of NTK eigenvectors during EoS in detail. Across different architectures, we observe that larger learning rates cause the leading eigenvectors of the final NTK, as well as the full NTK matrix, to have greater alignment with the training target. We then study the underlying mechanism of this phenomenon and provide a theoretical analysis for a two-layer linear network. Our study enhances the understanding of GD training dynamics in deep learning.
☆ MVA 2025 Small Multi-Object Tracking for Spotting Birds Challenge: Dataset, Methods, and Results
Small Multi-Object Tracking (SMOT) is particularly challenging when targets occupy only a few dozen pixels, rendering detection and appearance-based association unreliable. Building on the success of the MVA2023 SOD4SB challenge, this paper introduces the SMOT4SB challenge, which leverages temporal information to address limitations of single-frame detection. Our three main contributions are: (1) the SMOT4SB dataset, consisting of 211 UAV video sequences with 108,192 annotated frames under diverse real-world conditions, designed to capture motion entanglement where both camera and targets move freely in 3D; (2) SO-HOTA, a novel metric combining Dot Distance with HOTA to mitigate the sensitivity of IoU-based metrics to small displacements; and (3) a competitive MVA2025 challenge with 78 participants and 308 submissions, where the winning method achieved a 5.1x improvement over the baseline. This work lays a foundation for advancing SMOT in UAV scenarios with applications in bird strike avoidance, agriculture, fisheries, and ecological monitoring.
comment: This paper is the official challenge report for SMOT4SB and is published in the proceedings of MVA 2025 (19th International Conference on Machine Vision and Applications). Official challenge page: https://www.mva-org.jp/mva2025/challenge
☆ Autoregressive Speech Enhancement via Acoustic Tokens
In speech processing pipelines, improving the quality and intelligibility of real-world recordings is crucial. While supervised regression is the primary method for speech enhancement, audio tokenization is emerging as a promising alternative for a smooth integration with other modalities. However, research on speech enhancement using discrete representations is still limited. Previous work has mainly focused on semantic tokens, which tend to discard key acoustic details such as speaker identity. Additionally, these studies typically employ non-autoregressive models, assuming conditional independence of outputs and overlooking the potential improvements offered by autoregressive modeling. To address these gaps we: 1) conduct a comprehensive study of the performance of acoustic tokens for speech enhancement, including the effect of bitrate and noise strength; 2) introduce a novel transducer-based autoregressive architecture specifically designed for this task. Experiments on VoiceBank and Libri1Mix datasets show that acoustic tokens outperform semantic tokens in terms of preserving speaker identity, and that our autoregressive approach can further improve performance. Nevertheless, we observe that discrete representations still fall short compared to continuous ones, highlighting the need for further research in this area.
comment: 5 pages, 2 figures
☆ Assessing adaptive world models in machines with novel games
Human intelligence exhibits a remarkable capacity for rapid adaptation and effective problem-solving in novel and unfamiliar contexts. We argue that this profound adaptability is fundamentally linked to the efficient construction and refinement of internal representations of the environment, commonly referred to as world models, and we refer to this adaptation mechanism as world model induction. However, current understanding and evaluation of world models in artificial intelligence (AI) remains narrow, often focusing on static representations learned from training on a massive corpora of data, instead of the efficiency and efficacy of models in learning these representations through interaction and exploration within a novel environment. In this Perspective, we provide a view of world model induction drawing on decades of research in cognitive science on how humans learn and adapt so efficiently; we then call for a new evaluation framework for assessing adaptive world models in AI. Concretely, we propose a new benchmarking paradigm based on suites of carefully designed games with genuine, deep and continually refreshing novelty in the underlying game structures -- we refer to this kind of games as novel games. We detail key desiderata for constructing these games and propose appropriate metrics to explicitly challenge and evaluate the agent's ability for rapid world model induction. We hope that this new evaluation framework will inspire future evaluation efforts on world models in AI and provide a crucial step towards developing AI systems capable of the human-like rapid adaptation and robust generalization -- a critical component of artificial general intelligence.
comment: 17 pages, 4 figures
☆ Self Balancing Neural Network: A Novel Method to Estimate Average Treatment Effect
In observational studies, confounding variables affect both treatment and outcome. Moreover, instrumental variables also influence the treatment assignment mechanism. This situation sets the study apart from a standard randomized controlled trial, where the treatment assignment is random. Due to this situation, the estimated average treatment effect becomes biased. To address this issue, a standard approach is to incorporate the estimated propensity score when estimating the average treatment effect. However, these methods incur the risk of misspecification in propensity score models. To solve this issue, a novel method called the "Self balancing neural network" (Sbnet), which lets the model itself obtain its pseudo propensity score from the balancing net, is proposed in this study. The proposed method estimates the average treatment effect by using the balancing net as a key part of the feedforward neural network. This formulation resolves the estimation of the average treatment effect in one step. Moreover, the multi-pseudo propensity score framework, which is estimated from the diversified balancing net and used for the estimation of the average treatment effect, is presented. Finally, the proposed methods are compared with state-of-the-art methods on three simulation setups and real-world datasets. It has been shown that the proposed self-balancing neural network shows better performance than state-of-the-art methods.
☆ From Novelty to Imitation: Self-Distilled Rewards for Offline Reinforcement Learning
Offline Reinforcement Learning (RL) aims to learn effective policies from a static dataset without requiring further agent-environment interactions. However, its practical adoption is often hindered by the need for explicit reward annotations, which can be costly to engineer or difficult to obtain retrospectively. To address this, we propose ReLOAD (Reinforcement Learning with Offline Reward Annotation via Distillation), a novel reward annotation framework for offline RL. Unlike existing methods that depend on complex alignment procedures, our approach adapts Random Network Distillation (RND) to generate intrinsic rewards from expert demonstrations using a simple yet effective embedding discrepancy measure. First, we train a predictor network to mimic a fixed target network's embeddings based on expert state transitions. Later, the prediction error between these networks serves as a reward signal for each transition in the static dataset. This mechanism provides a structured reward signal without requiring handcrafted reward annotations. We provide a formal theoretical construct that offers insights into how RND prediction errors effectively serve as intrinsic rewards by distinguishing expert-like transitions. Experiments on the D4RL benchmark demonstrate that ReLOAD enables robust offline policy learning and achieves performance competitive with traditional reward-annotated methods.
☆ RONOM: Reduced-Order Neural Operator Modeling
Time-dependent partial differential equations are ubiquitous in physics-based modeling, but they remain computationally intensive in many-query scenarios, such as real-time forecasting, optimal control, and uncertainty quantification. Reduced-order modeling (ROM) addresses these challenges by constructing a low-dimensional surrogate model but relies on a fixed discretization, which limits flexibility across varying meshes during evaluation. Operator learning approaches, such as neural operators, offer an alternative by parameterizing mappings between infinite-dimensional function spaces, enabling adaptation to data across different resolutions. Whereas ROM provides rigorous numerical error estimates, neural operator learning largely focuses on discretization convergence and invariance without quantifying the error between the infinite-dimensional and the discretized operators. This work introduces the reduced-order neural operator modeling (RONOM) framework, which bridges concepts from ROM and operator learning. We establish a discretization error bound analogous to those in ROM, and get insights into RONOM's discretization convergence and discretization robustness. Moreover, two numerical examples are presented that compare RONOM to existing neural operators for solving partial differential equations. The results demonstrate that RONOM using standard vector-to-vector neural networks achieves comparable performance in input generalization and superior performance in both spatial super-resolution and discretization robustness, while also offering novel insights into temporal super-resolution scenarios.
☆ Large Language Models' Internal Perception of Symbolic Music
Large language models (LLMs) excel at modeling relationships between strings in natural language and have shown promise in extending to other symbolic domains like coding or mathematics. However, the extent to which they implicitly model symbolic music remains underexplored. This paper investigates how LLMs represent musical concepts by generating symbolic music data from textual prompts describing combinations of genres and styles, and evaluating their utility through recognition and generation tasks. We produce a dataset of LLM-generated MIDI files without relying on explicit musical training. We then train neural networks entirely on this LLM-generated MIDI dataset and perform genre and style classification as well as melody completion, benchmarking their performance against established models. Our results demonstrate that LLMs can infer rudimentary musical structures and temporal relationships from text, highlighting both their potential to implicitly encode musical patterns and their limitations due to a lack of explicit musical context, shedding light on their generative capabilities for symbolic music.
☆ PMKLC: Parallel Multi-Knowledge Learning-based Lossless Compression for Large-Scale Genomics Database KDD-25
Learning-based lossless compressors play a crucial role in large-scale genomic database backup, storage, transmission, and management. However, their 1) inadequate compression ratio, 2) low compression \& decompression throughput, and 3) poor compression robustness limit their widespread adoption and application in both industry and academia. To solve those challenges, we propose a novel \underline{P}arallel \underline{M}ulti-\underline{K}nowledge \underline{L}earning-based \underline{C}ompressor (PMKLC) with four crucial designs: 1) We propose an automated multi-knowledge learning-based compression framework as compressors' backbone to enhance compression ratio and robustness; 2) we design a GPU-accelerated ($s$,$k$)-mer encoder to optimize compression throughput and computing resource usage; 3) we introduce data block partitioning and Step-wise Model Passing (SMP) mechanisms for parallel acceleration; 4) We design two compression modes PMKLC-S and PMKLC-M to meet the complex application scenarios, where the former runs on a resource-constrained single GPU and the latter is multi-GPU accelerated. We benchmark PMKLC-S/M and 14 baselines (7 traditional and 7 leaning-based) on 15 real-world datasets with different species and data sizes. Compared to baselines on the testing datasets, PMKLC-S/M achieve the average compression ratio improvement up to 73.609\% and 73.480\%, the average throughput improvement up to 3.036$\times$ and 10.710$\times$, respectively. Besides, PMKLC-S/M also achieve the best robustness and competitive memory cost, indicating its greater stability against datasets with different probability distribution perturbations, and its strong ability to run on memory-constrained devices.
comment: Accepted via KDD-25
☆ FLDmamba: Integrating Fourier and Laplace Transform Decomposition with Mamba for Enhanced Time Series Prediction
Time series prediction, a crucial task across various domains, faces significant challenges due to the inherent complexities of time series data, including non-stationarity, multi-scale periodicity, and transient dynamics, particularly when tackling long-term predictions. While Transformer-based architectures have shown promise, their quadratic complexity with sequence length hinders their efficiency for long-term predictions. Recent advancements in State-Space Models, such as Mamba, offer a more efficient alternative for long-term modeling, but they cannot capture multi-scale periodicity and transient dynamics effectively. Meanwhile, they are susceptible to data noise issues in time series. This paper proposes a novel framework, FLDmamba (Fourier and Laplace Transform Decomposition Mamba), addressing these limitations. FLDmamba leverages the strengths of both Fourier and Laplace transforms to effectively capture both multi-scale periodicity, transient dynamics within time series data, and improve the robustness of the model to the data noise issue. Our extensive experiments demonstrate that FLDmamba achieves superior performance on time series prediction benchmarks, outperforming both Transformer-based and other Mamba-based architectures. To promote the reproducibility of our method, we have made both the code and data accessible via the following URL:{\href{https://github.com/AI4Science-WestlakeU/FLDmamba}{https://github.com/AI4Science-WestlakeU/\model}.
comment: 12 pages
☆ Multi-Channel Graph Neural Network for Financial Risk Prediction of NEEQ Enterprises
With the continuous evolution of China's multi-level capital market, the National Equities Exchange and Quotations (NEEQ), also known as the "New Third Board," has become a critical financing platform for small and medium-sized enterprises (SMEs). However, due to their limited scale and financial resilience, many NEEQ-listed companies face elevated risks of financial distress. To address this issue, we propose a multi-channel deep learning framework that integrates structured financial indicators, textual disclosures, and enterprise relationship data for comprehensive financial risk prediction. Specifically, we design a Triple-Channel Graph Isomorphism Network (GIN) that processes numeric, textual, and graph-based inputs separately. These modality-specific representations are fused using an attention-based mechanism followed by a gating unit to enhance robustness and prediction accuracy. Experimental results on data from 7,731 real-world NEEQ companies demonstrate that our model significantly outperforms traditional machine learning methods and single-modality baselines in terms of AUC, Precision, Recall, and F1 Score. This work provides theoretical and practical insights into risk modeling for SMEs and offers a data-driven tool to support financial regulators and investors.
comment: 10 pages, 4 figures. Submitted for conference review
☆ Compact Vision Transformer by Reduction of Kernel Complexity
Self-attention and transformer architectures have become foundational components in modern deep learning. Recent efforts have integrated transformer blocks into compact neural architectures for computer vision, giving rise to various efficient vision transformers. In this work, we introduce Transformer with Kernel Complexity Reduction, or KCR-Transformer, a compact transformer block equipped with differentiable channel selection, guided by a novel and sharp theoretical generalization bound. KCR-Transformer performs input/output channel selection in the MLP layers of transformer blocks to reduce the computational cost. Furthermore, we provide a rigorous theoretical analysis establishing a tight generalization bound for networks equipped with KCR-Transformer blocks. Leveraging such strong theoretical results, the channel pruning by KCR-Transformer is conducted in a generalization-aware manner, ensuring that the resulting network retains a provably small generalization error. Our KCR-Transformer is compatible with many popular and compact transformer networks, such as ViT and Swin, and it reduces the FLOPs of the vision transformers while maintaining or even improving the prediction accuracy. In the experiments, we replace all the transformer blocks in the vision transformers with KCR-Transformer blocks, leading to KCR-Transformer networks with different backbones. The resulting TCR-Transformers achieve superior performance on various computer vision tasks, achieving even better performance than the original models with even less FLOPs and parameters.
☆ A Comprehensive Survey of Electronic Health Record Modeling: From Deep Learning Approaches to Large Language Models
Artificial intelligence (AI) has demonstrated significant potential in transforming healthcare through the analysis and modeling of electronic health records (EHRs). However, the inherent heterogeneity, temporal irregularity, and domain-specific nature of EHR data present unique challenges that differ fundamentally from those in vision and natural language tasks. This survey offers a comprehensive overview of recent advancements at the intersection of deep learning, large language models (LLMs), and EHR modeling. We introduce a unified taxonomy that spans five key design dimensions: data-centric approaches, neural architecture design, learning-focused strategies, multimodal learning, and LLM-based modeling systems. Within each dimension, we review representative methods addressing data quality enhancement, structural and temporal representation, self-supervised learning, and integration with clinical knowledge. We further highlight emerging trends such as foundation models, LLM-driven clinical agents, and EHR-to-text translation for downstream reasoning. Finally, we discuss open challenges in benchmarking, explainability, clinical alignment, and generalization across diverse clinical settings. This survey aims to provide a structured roadmap for advancing AI-driven EHR modeling and clinical decision support. For a comprehensive list of EHR-related methods, kindly refer to https://survey-on-tabular-data.github.io/.
☆ Sample-Constrained Black Box Optimization for Audio Personalization AAAI 2024
We consider the problem of personalizing audio to maximize user experience. Briefly, we aim to find a filter $h^*$, which applied to any music or speech, will maximize the user's satisfaction. This is a black-box optimization problem since the user's satisfaction function is unknown. Substantive work has been done on this topic where the key idea is to play audio samples to the user, each shaped by a different filter $h_i$, and query the user for their satisfaction scores $f(h_i)$. A family of ``surrogate" functions is then designed to fit these scores and the optimization method gradually refines these functions to arrive at the filter $\hat{h}^*$ that maximizes satisfaction. In certain applications, we observe that a second type of querying is possible where users can tell us the individual elements $h^*[j]$ of the optimal filter $h^*$. Consider an analogy from cooking where the goal is to cook a recipe that maximizes user satisfaction. A user can be asked to score various cooked recipes (e.g., tofu fried rice) or to score individual ingredients (say, salt, sugar, rice, chicken, etc.). Given a budget of $B$ queries, where a query can be of either type, our goal is to find the recipe that will maximize this user's satisfaction. Our proposal builds on Sparse Gaussian Process Regression (GPR) and shows how a hybrid approach can outperform any one type of querying. Our results are validated through simulations and real world experiments, where volunteers gave feedback on music/speech audio and were able to achieve high satisfaction levels. We believe this idea of hybrid querying opens new problems in black-box optimization and solutions can benefit other applications beyond audio personalization.
comment: Published in AAAI 2024
☆ AnyPos: Automated Task-Agnostic Actions for Bimanual Manipulation
Vision-language-action (VLA) models have shown promise on task-conditioned control in complex settings such as bimanual manipulation. However, the heavy reliance on task-specific human demonstrations limits their generalization and incurs high data acquisition costs. In this work, we present a new notion of task-agnostic action paradigm that decouples action execution from task-specific conditioning, enhancing scalability, efficiency, and cost-effectiveness. To address the data collection challenges posed by this paradigm -- such as low coverage density, behavioral redundancy, and safety risks -- we introduce ATARA (Automated Task-Agnostic Random Actions), a scalable self-supervised framework that accelerates collection by over $ 30\times $ compared to human teleoperation. To further enable effective learning from task-agnostic data, which often suffers from distribution mismatch and irrelevant trajectories, we propose AnyPos, an inverse dynamics model equipped with Arm-Decoupled Estimation and a Direction-Aware Decoder (DAD). We additionally integrate a video-conditioned action validation module to verify the feasibility of learned policies across diverse manipulation tasks. Extensive experiments show that the AnyPos-ATARA pipeline yields a 51% improvement in test accuracy and achieves 30-40% higher success rates in downstream tasks such as lifting, pick-and-place, and clicking, using replay-based video validation. Project Page: https://embodiedfoundation.github.io/vidar_anypos
☆ Layer Separation Deep Learning Model with Auxiliary Variables for Partial Differential Equations
In this paper, we propose a new optimization framework, the layer separation (LySep) model, to improve the deep learning-based methods in solving partial differential equations. Due to the highly non-convex nature of the loss function in deep learning, existing optimization algorithms often converge to suboptimal local minima or suffer from gradient explosion or vanishing, resulting in poor performance. To address these issues, we introduce auxiliary variables to separate the layers of deep neural networks. Specifically, the output and its derivatives of each layer are represented by auxiliary variables, effectively decomposing the deep architecture into a series of shallow architectures. New loss functions with auxiliary variables are established, in which only variables from two neighboring layers are coupled. Corresponding algorithms based on alternating directions are developed, where many variables can be updated optimally in closed forms. Moreover, we provide theoretical analyses demonstrating the consistency between the LySep model and the original deep model. High-dimensional numerical results validate our theory and demonstrate the advantages of LySep in minimizing loss and reducing solution error.
☆ World Model-Based End-to-End Scene Generation for Accident Anticipation in Autonomous Driving
Reliable anticipation of traffic accidents is essential for advancing autonomous driving systems. However, this objective is limited by two fundamental challenges: the scarcity of diverse, high-quality training data and the frequent absence of crucial object-level cues due to environmental disruptions or sensor deficiencies. To tackle these issues, we propose a comprehensive framework combining generative scene augmentation with adaptive temporal reasoning. Specifically, we develop a video generation pipeline that utilizes a world model guided by domain-informed prompts to create high-resolution, statistically consistent driving scenarios, particularly enriching the coverage of edge cases and complex interactions. In parallel, we construct a dynamic prediction model that encodes spatio-temporal relationships through strengthened graph convolutions and dilated temporal operators, effectively addressing data incompleteness and transient visual noise. Furthermore, we release a new benchmark dataset designed to better capture diverse real-world driving risks. Extensive experiments on public and newly released datasets confirm that our framework enhances both the accuracy and lead time of accident anticipation, offering a robust solution to current data and modeling limitations in safety-critical autonomous driving applications.
☆ Domain-Enhanced Dual-Branch Model for Efficient and Interpretable Accident Anticipation
Developing precise and computationally efficient traffic accident anticipation system is crucial for contemporary autonomous driving technologies, enabling timely intervention and loss prevention. In this paper, we propose an accident anticipation framework employing a dual-branch architecture that effectively integrates visual information from dashcam videos with structured textual data derived from accident reports. Furthermore, we introduce a feature aggregation method that facilitates seamless integration of multimodal inputs through large models (GPT-4o, Long-CLIP), complemented by targeted prompt engineering strategies to produce actionable feedback and standardized accident archives. Comprehensive evaluations conducted on benchmark datasets (DAD, CCD, and A3D) validate the superior predictive accuracy, enhanced responsiveness, reduced computational overhead, and improved interpretability of our approach, thus establishing a new benchmark for state-of-the-art performance in traffic accident anticipation.
☆ Multimodal-Guided Dynamic Dataset Pruning for Robust and Efficient Data-Centric Learning
Modern deep models are trained on large real-world datasets, where data quality varies and redundancy is common. Data-centric approaches such as dataset pruning have shown promise in improving training efficiency and model performance. However, most existing methods rely on static heuristics or task-specific metrics, limiting their robustness and generalizability across domains. In this work, we introduce a dynamic dataset pruning framework that adaptively selects training samples based on both task-driven difficulty and cross-modality semantic consistency. By incorporating supervision from pretrained multimodal foundation models, our approach captures training dynamics while effectively filtering out uninformative samples. Our work highlights the potential of integrating cross-modality alignment for robust sample selection, advancing data-centric learning toward more efficient and robust practices across application domains.
☆ From SGD to Spectra: A Theory of Neural Network Weight Dynamics
Deep neural networks have revolutionized machine learning, yet their training dynamics remain theoretically unclear-we develop a continuous-time, matrix-valued stochastic differential equation (SDE) framework that rigorously connects the microscopic dynamics of SGD to the macroscopic evolution of singular-value spectra in weight matrices. We derive exact SDEs showing that squared singular values follow Dyson Brownian motion with eigenvalue repulsion, and characterize stationary distributions as gamma-type densities with power-law tails, providing the first theoretical explanation for the empirically observed 'bulk+tail' spectral structure in trained networks. Through controlled experiments on transformer and MLP architectures, we validate our theoretical predictions and demonstrate quantitative agreement between SDE-based forecasts and observed spectral evolution, providing a rigorous foundation for understanding why deep learning works.
♻ ☆ Federated Learning: A Survey on Privacy-Preserving Collaborative Intelligence
Federated Learning (FL) has emerged as a transformative paradigm in the field of distributed machine learning, enabling multiple clients such as mobile devices, edge nodes, or organizations to collaboratively train a shared global model without the need to centralize sensitive data. This decentralized approach addresses growing concerns around data privacy, security, and regulatory compliance, making it particularly attractive in domains such as healthcare, finance, and smart IoT systems. This survey provides a concise yet comprehensive overview of Federated Learning, beginning with its core architecture and communication protocol. We discuss the standard FL lifecycle, including local training, model aggregation, and global updates. A particular emphasis is placed on key technical challenges such as handling non-IID (non-independent and identically distributed) data, mitigating system and hardware heterogeneity, reducing communication overhead, and ensuring privacy through mechanisms like differential privacy and secure aggregation. Furthermore, we examine emerging trends in FL research, including personalized FL, cross-device versus cross-silo settings, and integration with other paradigms such as reinforcement learning and quantum computing. We also highlight real-world applications and summarize benchmark datasets and evaluation metrics commonly used in FL research. Finally, we outline open research problems and future directions to guide the development of scalable, efficient, and trustworthy FL systems.
♻ ☆ EgoVLA: Learning Vision-Language-Action Models from Egocentric Human Videos
Real robot data collection for imitation learning has led to significant advancements in robotic manipulation. However, the requirement for robot hardware in the process fundamentally constrains the scale of the data. In this paper, we explore training Vision-Language-Action (VLA) models using egocentric human videos. The benefit of using human videos is not only for their scale but more importantly for the richness of scenes and tasks. With a VLA trained on human video that predicts human wrist and hand actions, we can perform Inverse Kinematics and retargeting to convert the human actions to robot actions. We fine-tune the model using a few robot manipulation demonstrations to obtain the robot policy, namely EgoVLA. We propose a simulation benchmark called Ego Humanoid Manipulation Benchmark, where we design diverse bimanual manipulation tasks with demonstrations. We fine-tune and evaluate EgoVLA with Ego Humanoid Manipulation Benchmark and show significant improvements over baselines and ablate the importance of human data. Videos can be found on our website: https://rchalyang.github.io/EgoVLA
comment: More videos can be found on our website: https://rchalyang.github.io/EgoVLA
Learning from Reward-Free Offline Data: A Case for Planning with Latent Dynamics Models
A long-standing goal in AI is to build agents that can solve a variety of tasks across different environments, including previously unseen ones. Two dominant approaches tackle this challenge: (i) reinforcement learning (RL), which learns policies through trial and error, and (ii) optimal control, which plans actions using a learned or known dynamics model. However, their relative strengths and weaknesses remain underexplored in the setting where agents must learn from offline trajectories without reward annotations. In this work, we systematically analyze the performance of different RL and control-based methods under datasets of varying quality. On the RL side, we consider goal-conditioned and zero-shot approaches. On the control side, we train a latent dynamics model using the Joint Embedding Predictive Architecture (JEPA) and use it for planning. We study how dataset properties-such as data diversity, trajectory quality, and environment variability-affect the performance of these approaches. Our results show that model-free RL excels when abundant, high-quality data is available, while model-based planning excels in generalization to novel environment layouts, trajectory stitching, and data-efficiency. Notably, planning with a latent dynamics model emerges as a promising approach for zero-shot generalization from suboptimal data.
comment: Project web page: https://latent-planning.github.io/
♻ ☆ Retraining-Free Merging of Sparse MoE via Hierarchical Clustering
Sparse Mixture-of-Experts (SMoE) models represent a significant advancement in large language model (LLM) development through their efficient parameter utilization. These models achieve substantial performance improvements at reduced inference costs. However, the deployment of SMoE models faces constraints from extensive memory requirements of expert components in resource-limited environments. To address these limitations, this paper introduces Hierarchical Clustering for Sparsely activated Mixture of Experts (HC-SMoE), a task-agnostic expert merging framework for parameter reduction without retraining. HC-SMoE introduces a novel hierarchical clustering approach based on expert outputs to ensure merging robustness independent of routing decisions. The proposed output-based clustering method enables effective capture of functional relationships between experts for large-scale architectures. We provide theoretical analysis and comprehensive evaluations across multiple zero-shot language tasks to demonstrate HC-SMoE's effectiveness in state-of-the-art models including Qwen and Mixtral. The experimental results validate HC-SMoE's superior performance and practical applicability for real-world deployments.
comment: Code: https://github.com/wazenmai/HC-SMoE
♻ ☆ Advancing Seasonal Prediction of Tropical Cyclone Activity with a Hybrid AI-Physics Climate Model
Machine learning (ML) models are successful with weather forecasting and have shown progress in climate simulations, yet leveraging them for useful climate predictions needs exploration. Here we show this feasibility using Neural General Circulation Model (NeuralGCM), a hybrid ML-physics atmospheric model developed by Google, for seasonal predictions of large-scale atmospheric variability and Northern Hemisphere tropical cyclone (TC) activity. Inspired by physical model studies, we simplify boundary conditions, assuming sea surface temperature (SST) and sea ice follow their climatological cycle but persist anomalies present at the initialization time. With such forcings, NeuralGCM can generate 100 simulation days in ~8 minutes with a single Graphics Processing Unit (GPU), while simulating realistic atmospheric circulation and TC climatology patterns. This configuration yields useful seasonal predictions (July to November) for the tropical atmosphere and various TC activity metrics. Notably, the predicted and observed TC frequency in the North Atlantic and East Pacific basins are significantly correlated during 1990 to 2023 (r=~0.7), suggesting prediction skill comparable to existing physical GCMs. Despite challenges associated with model resolution and simplified boundary forcings, the model-predicted interannual variations demonstrate significant correlations with the observation, including the sub-basin TC tracks (p<0.1) and basin-wide accumulated cyclone energy (p<0.01) of the North Atlantic and North Pacific basins. These findings highlight the promise of leveraging ML models with physical insights to model TC risks and deliver seamless weather-climate predictions.
♻ ☆ SIDDA: SInkhorn Dynamic Domain Adaptation for Image Classification with Equivariant Neural Networks
Modern neural networks (NNs) often do not generalize well in the presence of a "covariate shift"; that is, in situations where the training and test data distributions differ, but the conditional distribution of classification labels remains unchanged. In such cases, NN generalization can be reduced to a problem of learning more domain-invariant features. Domain adaptation (DA) methods include a range of techniques aimed at achieving this; however, these methods have struggled with the need for extensive hyperparameter tuning, which then incurs significant computational costs. In this work, we introduce SIDDA, an out-of-the-box DA training algorithm built upon the Sinkhorn divergence, that can achieve effective domain alignment with minimal hyperparameter tuning and computational overhead. We demonstrate the efficacy of our method on multiple simulated and real datasets of varying complexity, including simple shapes, handwritten digits, and real astronomical observations. SIDDA is compatible with a variety of NN architectures, and it works particularly well in improving classification accuracy and model calibration when paired with equivariant neural networks (ENNs). We find that SIDDA enhances the generalization capabilities of NNs, achieving up to a $\approx40\%$ improvement in classification accuracy on unlabeled target data. We also study the efficacy of DA on ENNs with respect to the varying group orders of the dihedral group $D_N$, and find that the model performance improves as the degree of equivariance increases. Finally, we find that SIDDA enhances model calibration on both source and target data--achieving over an order of magnitude improvement in the ECE and Brier score. SIDDA's versatility, combined with its automated approach to domain alignment, has the potential to advance multi-dataset studies by enabling the development of highly generalizable models.
comment: 25 pages, 5 figures, 4 tables. code available at: https://github.com/deepskies/SIDDA
♻ ☆ crowd-hpo: Realistic Hyperparameter Optimization and Benchmarking for Learning from Crowds with Noisy Labels
Crowdworking is a cost-efficient solution for acquiring class labels. Since these labels are subject to noise, various approaches to learning from crowds have been proposed. Typically, these approaches are evaluated with default hyperparameter configurations, resulting in unfair and suboptimal performance, or with hyperparameter configurations tuned via a validation set with ground truth class labels, representing an often unrealistic scenario. Moreover, both setups can produce different approach rankings, complicating study comparisons. Therefore, we introduce crowd-hpo as a framework for evaluating approaches to learning from crowds in combination with criteria to select well-performing hyperparameter configurations with access only to noisy crowd-labeled validation data. Extensive experiments with neural networks demonstrate that these criteria select hyperparameter configurations, which improve the learning from crowd approaches' generalization performances, measured on separate test sets with ground truth labels. Hence, incorporating such criteria into experimental studies is essential for enabling fairer and more realistic benchmarking.
comment: Under review
♻ ☆ Generative Diffusion Models for Resource Allocation in Wireless Networks
This paper proposes a supervised training algorithm for learning stochastic resource allocation policies with generative diffusion models (GDMs). We formulate the allocation problem as the maximization of an ergodic utility function subject to ergodic Quality of Service (QoS) constraints. Given samples from a stochastic expert policy that yields a near-optimal solution to the constrained optimization problem, we train a GDM policy to imitate the expert and generate new samples from the optimal distribution. We achieve near-optimal performance through the sequential execution of the generated samples. To enable generalization to a family of network configurations, we parameterize the backward diffusion process with a graph neural network (GNN) architecture. We present numerical results in a case study of power control.
♻ ☆ Do you know what q-means?
Clustering is one of the most important tools for analysis of large datasets, and perhaps the most popular clustering algorithm is Lloyd's algorithm for $k$-means. This algorithm takes $n$ vectors $V=[v_1,\dots,v_n]\in\mathbb{R}^{d\times n}$ and outputs $k$ centroids $c_1,\dots,c_k\in\mathbb{R}^d$; these partition the vectors into clusters based on which centroid is closest to a particular vector. We present a classical $\varepsilon$-$k$-means algorithm that performs an approximate version of one iteration of Lloyd's algorithm with time complexity $\tilde{O}\big(\frac{\|V\|_F^2}{n}\frac{k^{2}d}{\varepsilon^2}(k + \log{n})\big)$, exponentially improving the dependence on the data size $n$ and matching that of the "$q$-means" quantum algorithm originally proposed by Kerenidis, Landman, Luongo, and Prakash (NeurIPS'19). Moreover, we propose an improved $q$-means quantum algorithm with time complexity $\tilde{O}\big(\frac{\|V\|_F}{\sqrt{n}}\frac{k^{3/2}d}{\varepsilon}(\sqrt{k}+\sqrt{d})(\sqrt{k} + \log{n})\big)$ that quadratically improves the runtime of our classical $\varepsilon$-$k$-means algorithm in several parameters. Our quantum algorithm does not rely on quantum linear algebra primitives of prior work, but instead only uses QRAM to prepare simple states based on the current iteration's clusters and multivariate quantum amplitude estimation. Finally, we provide classical and quantum query lower bounds, showing that our algorithms are optimal in most parameters.
comment: 21 pages. v2: improved the quantum complexity, references added; v3: new co-author added, new algorithms and upper bounds, improved old upper bounds, new lower bounds, references added
♻ ☆ A Roadmap for Climate-Relevant Robotics Research
Climate change is one of the defining challenges of the 21st century, and many in the robotics community are looking for ways to contribute. This paper presents a roadmap for climate-relevant robotics research, identifying high-impact opportunities for collaboration between roboticists and experts across climate domains such as energy, the built environment, transportation, industry, land use, and Earth sciences. These applications include problems such as energy systems optimization, construction, precision agriculture, building envelope retrofits, autonomous trucking, and large-scale environmental monitoring. Critically, we include opportunities to apply not only physical robots but also the broader robotics toolkit - including planning, perception, control, and estimation algorithms - to climate-relevant problems. A central goal of this roadmap is to inspire new research directions and collaboration by highlighting specific, actionable problems at the intersection of robotics and climate. This work represents a collaboration between robotics researchers and domain experts in various climate disciplines, and it serves as an invitation to the robotics community to bring their expertise to bear on urgent climate priorities.
♻ ☆ Approximation Rates for Shallow ReLU$^k$ Neural Networks on Sobolev Spaces via the Radon Transform
Let $\Omega\subset \mathbb{R}^d$ be a bounded domain. We consider the problem of how efficiently shallow neural networks with the ReLU$^k$ activation function can approximate functions from Sobolev spaces $W^s(L_p(\Omega))$ with error measured in the $L_q(\Omega)$-norm. Utilizing the Radon transform and recent results from discrepancy theory, we provide a simple proof of nearly optimal approximation rates in a variety of cases, including when $q\leq p$, $p\geq 2$, and $s \leq k + (d+1)/2$. The rates we derive are optimal up to logarithmic factors, and significantly generalize existing results. An interesting consequence is that the adaptivity of shallow ReLU$^k$ neural networks enables them to obtain optimal approximation rates for smoothness up to order $s = k + (d+1)/2$, even though they represent piecewise polynomials of fixed degree $k$.
♻ ☆ VectorFit : Adaptive Singular & Bias Vector Fine-Tuning of Pre-trained Foundation Models
Popular PEFT methods reduce trainable parameter count for fine-tuning by parameterizing new low-rank or sparse trainable weights in parallel to the frozen pre-trained weights $W$. However, these weights are trained from scratch, and there exists a performance gap between these methods and full fine-tuning, especially in low-budget settings. We introduce VectorFit, a new way of parameterization that efficiently utilizes the existing knowledge embedded in $W$ by adaptively training their singular vectors and biases. We show that utilizing the structural and transformational properties of $W$ in this way can lead to high-rank incremental weight matrices $\Delta W$, comparable to that of full fine-tuning. VectorFit delivers superior results with \textbf{9$\boldsymbol\times$} fewer trainable parameters than the leading PEFT methods. Through comprehensive experiments across 19 datasets covering a wide range of language and vision tasks such as natural language understanding and generation, question answering, image classification, and image generation, we demonstrate that VectorFit surpasses baselines in terms of performance as a function of parameter-efficiency.
♻ ☆ Multiple-Frequencies Population-Based Training
Reinforcement Learning's high sensitivity to hyperparameters is a source of instability and inefficiency, creating significant challenges for practitioners. Hyperparameter Optimization (HPO) algorithms have been developed to address this issue, among them Population-Based Training (PBT) stands out for its ability to generate hyperparameters schedules instead of fixed configurations. PBT trains a population of agents, each with its own hyperparameters, frequently ranking them and replacing the worst performers with mutations of the best agents. These intermediate selection steps can cause PBT to focus on short-term improvements, leading it to get stuck in local optima and eventually fall behind vanilla Random Search over longer timescales. This paper studies how this greediness issue is connected to the choice of evolution frequency, the rate at which the selection is done. We propose Multiple-Frequencies Population-Based Training (MF-PBT), a novel HPO algorithm that addresses greediness by employing sub-populations, each evolving at distinct frequencies. MF-PBT introduces a migration process to transfer information between sub-populations, with an asymmetric design to balance short and long-term optimization. Extensive experiments on the Brax suite demonstrate that MF-PBT improves sample efficiency and long-term performance, even without actually tuning hyperparameters.
comment: RLC25 - Camera-ready
♻ ☆ V-Max: A Reinforcement Learning Framework for Autonomous Driving
Learning-based decision-making has the potential to enable generalizable Autonomous Driving (AD) policies, reducing the engineering overhead of rule-based approaches. Imitation Learning (IL) remains the dominant paradigm, benefiting from large-scale human demonstration datasets, but it suffers from inherent limitations such as distribution shift and imitation gaps. Reinforcement Learning (RL) presents a promising alternative, yet its adoption in AD remains limited due to the lack of standardized and efficient research frameworks. To this end, we introduce V-Max, an open research framework providing all the necessary tools to make RL practical for AD. V-Max is built on Waymax, a hardware-accelerated AD simulator designed for large-scale experimentation. We extend it using ScenarioNet's approach, enabling the fast simulation of diverse AD datasets.
comment: RLC 25 - Camera-ready
♻ ☆ Compositional Discrete Latent Code for High Fidelity, Productive Diffusion Models
We argue that diffusion models' success in modeling complex distributions is, for the most part, coming from their input conditioning. This paper investigates the representation used to condition diffusion models from the perspective that ideal representations should improve sample fidelity, be easy to generate, and be compositional to allow out-of-training samples generation. We introduce Discrete Latent Code (DLC), an image representation derived from Simplicial Embeddings trained with a self-supervised learning objective. DLCs are sequences of discrete tokens, as opposed to the standard continuous image embeddings. They are easy to generate and their compositionality enables sampling of novel images beyond the training distribution. Diffusion models trained with DLCs have improved generation fidelity, establishing a new state-of-the-art for unconditional image generation on ImageNet. Additionally, we show that composing DLCs allows the image generator to produce out-of-distribution samples that coherently combine the semantics of images in diverse ways. Finally, we showcase how DLCs can enable text-to-image generation by leveraging large-scale pretrained language models. We efficiently finetune a text diffusion language model to generate DLCs that produce novel samples outside of the image generator training distribution.
comment: In submission, 22 pages, 7 tables, 12 figures
♻ ☆ Branching Stein Variational Gradient Descent for sampling multimodal distributions
We propose a novel particle-based variational inference method designed to work with multimodal distributions. Our approach, referred to as Branched Stein Variational Gradient Descent (BSVGD), extends the classical Stein Variational Gradient Descent (SVGD) algorithm by incorporating a random branching mechanism that encourages the exploration of the state space. In this work, a theoretical guarantee for the convergence in distribution is presented, as well as numerical experiments to validate the suitability of our algorithm. Performance comparisons between the BSVGD and the SVGD are presented using the Wasserstein distance between samples and the corresponding computational times.
♻ ☆ Recent Advances in Simulation-based Inference for Gravitational Wave Data Analysis
The detection of gravitational waves by the LIGO-Virgo-KAGRA collaboration has ushered in a new era of observational astronomy, emphasizing the need for rapid and detailed parameter estimation and population-level analyses. Traditional Bayesian inference methods, particularly Markov chain Monte Carlo, face significant computational challenges when dealing with the high-dimensional parameter spaces and complex noise characteristics inherent in gravitational wave data. This review examines the emerging role of simulation-based inference methods in gravitational wave astronomy, with a focus on approaches that leverage machine-learning techniques such as normalizing flows and neural posterior estimation. We provide a comprehensive overview of the theoretical foundations underlying various simulation-based inference methods, including neural posterior estimation, neural ratio estimation, neural likelihood estimation, flow matching, and consistency models. We explore the applications of these methods across diverse gravitational wave data processing scenarios, from single-source parameter estimation and overlapping signal analysis to testing general relativity and conducting population studies. Although these techniques demonstrate speed improvements over traditional methods in controlled studies, their model-dependent nature and sensitivity to prior assumptions are barriers to their widespread adoption. Their accuracy, which is similar to that of conventional methods, requires further validation across broader parameter spaces and noise conditions.
comment: 30 pages, 6 figures, 1 table. Minor clarifications added on page 3. Literature covered up to early 2025
♻ ☆ Bounding the Worst-class Error: A Boosting Approach IJCNN2025
This paper tackles the problem of the worst-class error rate, instead of the standard error rate averaged over all classes. For example, a three-class classification task with class-wise error rates of 10%, 10%, and 40% has a worst-class error rate of 40%, whereas the average is 20% under the class-balanced condition. The worst-class error is important in many applications. For example, in a medical image classification task, it would not be acceptable for the malignant tumor class to have a 40% error rate, while the benign and healthy classes have a 10% error rates. To avoid overfitting in worst-class error minimization using Deep Neural Networks (DNNs), we design a problem formulation for bounding the worst-class error instead of achieving zero worst-class error. Moreover, to correctly bound the worst-class error, we propose a boosting approach which ensembles DNNs. We give training and generalization worst-class-error bound. Experimental results show that the algorithm lowers worst-class test error rates while avoiding overfitting to the training set. This code is available at https://github.com/saito-yuya/Bounding-the-Worst-class-error-A-Boosting-Approach.
comment: Accepted at IJCNN2025
♻ ☆ PINT: Physics-Informed Neural Time Series Models with Applications to Long-term Inference on WeatherBench 2m-Temperature Data
This paper introduces PINT (Physics-Informed Neural Time Series Models), a framework that integrates physical constraints into neural time series models to improve their ability to capture complex dynamics. We apply PINT to the ERA5 WeatherBench dataset, focusing on long-term forecasting of 2m-temperature data. PINT incorporates the Simple Harmonic Oscillator Equation as a physics-informed prior, embedding its periodic dynamics into RNN, LSTM, and GRU architectures. This equation's analytical solutions (sine and cosine functions) facilitate rigorous evaluation of the benefits of incorporating physics-informed constraints. By benchmarking against a linear regression baseline derived from its exact solutions, we quantify the impact of embedding physical principles in data-driven models. Unlike traditional time series models that rely on future observations, PINT is designed for practical forecasting. Using only the first 90 days of observed data, it iteratively predicts the next two years, addressing challenges posed by limited real-time updates. Experiments on the WeatherBench dataset demonstrate PINT's ability to generalize, capture periodic trends, and align with physical principles. This study highlights the potential of physics-informed neural models in bridging machine learning and interpretable climate applications. Our models and datasets are publicly available on GitHub: https://github.com/KV-Park.
♻ ☆ Task-Circuit Quantization: Leveraging Knowledge Localization and Interpretability for Compression
Post-training quantization (PTQ) reduces a model's memory footprint by mapping full precision weights into low bit weights without costly retraining, but can degrade its downstream performance especially in low 2- to 3-bit settings. We develop a new mixed-precision PTQ approach, Task-Circuit Quantization (TaCQ), that draws parallels to automated circuit discovery, directly conditioning the quantization process on specific weight circuits -- which we define as sets of weights associated with downstream task performance. These weights are kept as 16-bit weights, while others are quantized, maintaining performance while only adding a marginal memory cost. Specifically, TaCQ contrasts unquantized model weights with a uniformly-quantized model to estimate the expected change in weights due to quantization and uses gradient information to predict the resulting impact on task performance, allowing us to preserve task-specific weights. We compare TaCQ-based quantization to existing mixed-precision quantization methods when conditioning both on general-purpose and task-specific data. Across QA, math reasoning, and text-to-SQL tasks for both Llama-3 and Qwen2.5, we find that TaCQ outperforms baselines using the same calibration data and a lower weight budget, achieving major improvements in the 2 and 3-bit regime. With only 3.1 bits we are able to recover 96% of Llama-3-8B-Instruct's unquantized 16-bit MMLU performance, obtaining a 5.25% absolute improvement over SPQR. We also observe consistently large gains over existing methods in the 2-bit regime, with an average gain of 14.74% over the strongest baseline, SliM-LLM. Moreover, we observe a 7.20% gain without conditioning on specific tasks, showing TaCQ's ability to identify important weights is not limited to task-conditioned settings.
comment: COLM 2025 Camera Ready. Code: https://github.com/The-Inscrutable-X/TACQ
♻ ☆ Unified Triplet-Level Hallucination Evaluation for Large Vision-Language Models
Despite the outstanding performance in vision-language reasoning, Large Vision-Language Models (LVLMs) might generate hallucinated contents that do not exist in the given image. Most existing LVLM hallucination benchmarks are constrained to evaluate the object-related hallucinations. However, the potential hallucination on the relations between two objects, i.e., relation hallucination, still lacks investigation. To remedy that, we design a unified framework to measure the object and relation hallucination in LVLMs simultaneously. The core idea of our framework is to evaluate hallucinations via (object, relation, object) triplets extracted from LVLMs' responses, making it easily generalizable to different vision-language tasks. Based on our framework, we further introduce Tri-HE, a novel Triplet-level Hallucination Evaluation benchmark which can be used to study both object and relation hallucination at the same time. With comprehensive evaluations on Tri-HE, we observe that the relation hallucination issue is even more serious than object hallucination among existing LVLMs, highlighting a previously neglected problem towards reliable LVLMs. Moreover, based on our findings, we design a simple training-free approach that effectively mitigates hallucinations for LVLMs. Our dataset and code for the reproduction of our experiments are available publicly at https://github.com/wujunjie1998/Tri-HE.
comment: Accepted by TMLR 2025. Project Page: https://kaichen1998.github.io/projects/tri-he/
♻ ☆ Uni-Instruct: One-step Diffusion Model through Unified Diffusion Divergence Instruction
In this paper, we unify more than 10 existing one-step diffusion distillation approaches, such as Diff-Instruct, DMD, SIM, SiD, $f$-distill, etc, inside a theory-driven framework which we name the \textbf{\emph{Uni-Instruct}}. Uni-Instruct is motivated by our proposed diffusion expansion theory of the $f$-divergence family. Then we introduce key theories that overcome the intractability issue of the original expanded $f$-divergence, resulting in an equivalent yet tractable loss that effectively trains one-step diffusion models by minimizing the expanded $f$-divergence family. The novel unification introduced by Uni-Instruct not only offers new theoretical contributions that help understand existing approaches from a high-level perspective but also leads to state-of-the-art one-step diffusion generation performances. On the CIFAR10 generation benchmark, Uni-Instruct achieves record-breaking Frechet Inception Distance (FID) values of \textbf{\emph{1.46}} for unconditional generation and \textbf{\emph{1.38}} for conditional generation. On the ImageNet-$64\times 64$ generation benchmark, Uni-Instruct achieves a new SoTA one-step generation FID of \textbf{\emph{1.02}}, which outperforms its 79-step teacher diffusion with a significant improvement margin of 1.33 (1.02 vs 2.35). We also apply Uni-Instruct on broader tasks like text-to-3D generation. For text-to-3D generation, Uni-Instruct gives decent results, which slightly outperforms previous methods, such as SDS and VSD, in terms of both generation quality and diversity. Both the solid theoretical and empirical contributions of Uni-Instruct will potentially help future studies on one-step diffusion distillation and knowledge transferring of diffusion models.
♻ ☆ Truthful Elicitation of Imprecise Forecasts UAI 2025
The quality of probabilistic forecasts is crucial for decision-making under uncertainty. While proper scoring rules incentivize truthful reporting of precise forecasts, they fall short when forecasters face epistemic uncertainty about their beliefs, limiting their use in safety-critical domains where decision-makers (DMs) prioritize proper uncertainty management. To address this, we propose a framework for scoring imprecise forecasts -- forecasts given as a set of beliefs. Despite existing impossibility results for deterministic scoring rules, we enable truthful elicitation by drawing connection to social choice theory and introducing a two-way communication framework where DMs first share their aggregation rules (e.g., averaging or min-max) used in downstream decisions for resolving forecast ambiguity. This, in turn, helps forecasters resolve indecision during elicitation. We further show that truthful elicitation of imprecise forecasts is achievable using proper scoring rules randomized over the aggregation procedure. Our approach allows DM to elicit and integrate the forecaster's epistemic uncertainty into their decision-making process, thus improving credibility.
comment: Accepted at UAI 2025 for Oral Presentation (fixed formatting)
♻ ☆ Super Resolution for Renewable Energy Resource Data With Wind From Reanalysis Data and Application to Ukraine
With a potentially increasing share of the electricity grid relying on wind to provide generating capacity and energy, there is an expanding global need for historically accurate, spatiotemporally continuous, high-resolution wind data. Conventional downscaling methods for generating these data based on numerical weather prediction have a high computational burden and require extensive tuning for historical accuracy. In this work, we present a novel deep learning-based spatiotemporal downscaling method using generative adversarial networks (GANs) for generating historically accurate high-resolution wind resource data from the European Centre for Medium-Range Weather Forecasting Reanalysis version 5 data (ERA5). In contrast to previous approaches, which used coarsened high-resolution data as low-resolution training data, we use true low-resolution simulation outputs. We show that by training a GAN model with ERA5 as the low-resolution input and Wind Integration National Dataset Toolkit (WTK) data as the high-resolution target, we achieved results comparable in historical accuracy and spatiotemporal variability to conventional dynamical downscaling. This GAN-based downscaling method additionally reduces computational costs over dynamical downscaling by two orders of magnitude. We applied this approach to downscale 30 km, hourly ERA5 data to 2 km, 5 min wind data for January 2000 through December 2023 at multiple hub heights over Ukraine, Moldova, and part of Romania. This 24-year data record is the first member of the super-resolution for renewable energy resource data with wind from the reanalysis data dataset (Sup3rWind).
comment: 22 pages, 9 figures
♻ ☆ On the Effectiveness of the z-Transform Method in Quadratic Optimization
The z-transform of a sequence is a classical tool used within signal processing, control theory, computer science, and electrical engineering. It allows for studying sequences from their generating functions, with many operations that can be equivalently defined on the original sequence and its $z$-transform. In particular, the z-transform method focuses on asymptotic behaviors and allows the use of Taylor expansions. We present a sequence of results of increasing significance and difficulty for linear models and optimization algorithms, demonstrating the effectiveness and versatility of the z-transform method in deriving new asymptotic results. Starting from the simplest gradient descent iterations in an infinite-dimensional Hilbert space, we show how the spectral dimension characterizes the convergence behavior. We then extend the analysis to Nesterov acceleration, averaging techniques, and stochastic gradient descent.
♻ ☆ MedPix 2.0: A Comprehensive Multimodal Biomedical Data set for Advanced AI Applications with Retrieval Augmented Generation and Knowledge Graphs
The increasing interest in developing Artificial Intelligence applications in the medical domain, suffers from the lack of high-quality data set, mainly due to privacy-related issues. In addition, the recent increase in Vision Language Models (VLM) leads to the need for multimodal medical data sets, where clinical reports and findings are attached to the corresponding medical scans. This paper illustrates the entire workflow for building the MedPix 2.0 data set. Starting with the well-known multimodal data set MedPix\textsuperscript{\textregistered}, mainly used by physicians, nurses, and healthcare students for Continuing Medical Education purposes, a semi-automatic pipeline was developed to extract visual and textual data followed by a manual curing procedure in which noisy samples were removed, thus creating a MongoDB database. Along with the data set, we developed a Graphical User Interface aimed at navigating efficiently the MongoDB instance and obtaining the raw data that can be easily used for training and/or fine-tuning VLMs. To enforce this point, in this work, we first recall DR-Minerva, a Retrieve Augmented Generation-based VLM model trained upon MedPix 2.0. DR-Minerva predicts the body part and the modality used to scan its input image. We also propose the extension of DR-Minerva with a Knowledge Graph that uses Llama 3.1 Instruct 8B, and leverages MedPix 2.0. The resulting architecture can be queried in a end-to-end manner, as a medical decision support system. MedPix 2.0 is available on GitHub.
♻ ☆ Uncertainty quantification for White Matter Hyperintensity segmentation detects silent failures and improves automated Fazekas quantification
White Matter Hyperintensities (WMH) are key neuroradiological markers of small vessel disease present in brain MRI. Assessment of WMH is important in research and clinics. However, WMH are challenging to segment due to their high variability in shape, location, size, poorly defined borders, and similar intensity profile to other pathologies (e.g stroke lesions) and artefacts (e.g head motion). In this work, we assess the utility and semantic properties of the most effective techniques for uncertainty quantification (UQ) in segmentation for the WMH segmentation task across multiple test-time data distributions. We find UQ techniques reduce 'silent failure' by identifying in UQ maps small WMH clusters in the deep white matter that are unsegmented by the model. A combination of Stochastic Segmentation Networks with Deep Ensembles also yields the highest Dice and lowest Absolute Volume Difference % (AVD) score and can highlight areas where there is ambiguity between WMH and stroke lesions. We further demonstrate the downstream utility of UQ, proposing a novel method for classification of the clinical Fazekas score using spatial features extracted from voxelwise WMH probability and UQ maps. We show that incorporating WMH uncertainty information improves Fazekas classification performance and calibration. Our model with (UQ and spatial WMH features)/(spatial WMH features)/(WMH volume only) achieves a balanced accuracy score of 0.74/0.67/0.62, and root brier score of 0.65/0.72/0.74 in the Deep WMH and balanced accuracy of 0.74/0.73/0.71 and root brier score of 0.64/0.66/0.68 in the Periventricular region. We further demonstrate that stochastic UQ techniques with high sample diversity can improve the detection of poor quality segmentations.
comment: 34 pages (or 19 not including appendix) 28 figures (or 10 not including appendix)
♻ ☆ The late-stage training dynamics of (stochastic) subgradient descent on homogeneous neural networks COLT 2025
We analyze the implicit bias of constant step stochastic subgradient descent (SGD). We consider the setting of binary classification with homogeneous neural networks - a large class of deep neural networks with ReLU-type activation functions such as MLPs and CNNs without biases. We interpret the dynamics of normalized SGD iterates as an Euler-like discretization of a conservative field flow that is naturally associated to the normalized classification margin. Owing to this interpretation, we show that normalized SGD iterates converge to the set of critical points of the normalized margin at late-stage training (i.e., assuming that the data is correctly classified with positive normalized margin). Up to our knowledge, this is the first extension of the analysis of Lyu and Li (2020) on the discrete dynamics of gradient descent to the nonsmooth and stochastic setting. Our main result applies to binary classification with exponential or logistic losses. We additionally discuss extensions to more general settings.
comment: Accepted/presented at the 38th Annual Conference on Learning Theory (COLT 2025)
♻ ☆ (Almost) Free Modality Stitching of Foundation Models
Foundation multi-modal models are often designed by stitching of multiple existing pretrained uni-modal models: for example, an image classifier with an text model. This stitching process is performed by training a connector module that aims to align the representation spaces of these uni-modal models towards a multi-modal objective. However, given the complexity of training such connectors on large scale web-based datasets coupled with the ever-increasing number of available pretrained uni-modal models, the task of uni-modal models selection and subsequent connector module training becomes computationally demanding. To address this under-studied critical problem, we propose Hypernetwork Model Alignment (Hyma), a novel all-in-one solution for optimal uni-modal model selection and connector training by leveraging hypernetworks. Specifically, our framework utilizes the parameter prediction capability of a hypernetwork to obtain jointly trained connector modules for $N \times M$ combinations of uni-modal models. In our experiments, Hyma reduces the cost of searching for the best performing uni-modal model pair by $10\times$, while matching the ranking and trained connector performance obtained via grid search across a suite of diverse multi-modal benchmarks.
comment: Pre-print
♻ ☆ Characterizing Dynamical Stability of Stochastic Gradient Descent in Overparameterized Learning
For overparameterized optimization tasks, such as those found in modern machine learning, global minima are generally not unique. In order to understand generalization in these settings, it is vital to study to which minimum an optimization algorithm converges. The possibility of having minima that are unstable under the dynamics imposed by the optimization algorithm limits the potential minima that the algorithm can find. In this paper, we characterize the global minima that are dynamically stable/unstable for both deterministic and stochastic gradient descent (SGD). In particular, we introduce a characteristic Lyapunov exponent that depends on the local dynamics around a global minimum and rigorously prove that the sign of this Lyapunov exponent determines whether SGD can accumulate at the respective global minimum.
♻ ☆ A Progressive Image Restoration Network for High-order Degradation Imaging in Remote Sensing
Recently, deep learning methods have gained remarkable achievements in the field of image restoration for remote sensing (RS). However, most existing RS image restoration methods focus mainly on conventional first-order degradation models, which may not effectively capture the imaging mechanisms of remote sensing images. Furthermore, many RS image restoration approaches that use deep learning are often criticized for their lacks of architecture transparency and model interpretability. To address these problems, we propose a novel progressive restoration network for high-order degradation imaging (HDI-PRNet), to progressively restore different image degradation. HDI-PRNet is developed based on the theoretical framework of degradation imaging, also Markov properties of the high-order degradation process and Maximum a posteriori (MAP) estimation, offering the benefit of mathematical interpretability within the unfolding network. The framework is composed of three main components: a module for image denoising that relies on proximal mapping prior learning, a module for image deblurring that integrates Neumann series expansion with dual-domain degradation learning, and a module for super-resolution. Extensive experiments demonstrate that our method achieves superior performance on both synthetic and real remote sensing images.
comment: 17 pages, Accepted to Transactions on Geoscience and Remote Sensing (TGRS), July 16, 2025
♻ ☆ A Brain Tumor Segmentation Method Based on CLIP and 3D U-Net with Cross-Modal Semantic Guidance and Multi-Level Feature Fusion
Precise segmentation of brain tumors from magnetic resonance imaging (MRI) is essential for neuro-oncology diagnosis and treatment planning. Despite advances in deep learning methods, automatic segmentation remains challenging due to tumor morphological heterogeneity and complex three-dimensional spatial relationships. Current techniques primarily rely on visual features extracted from MRI sequences while underutilizing semantic knowledge embedded in medical reports. This research presents a multi-level fusion architecture that integrates pixel-level, feature-level, and semantic-level information, facilitating comprehensive processing from low-level data to high-level concepts. The semantic-level fusion pathway combines the semantic understanding capabilities of Contrastive Language-Image Pre-training (CLIP) models with the spatial feature extraction advantages of 3D U-Net through three mechanisms: 3D-2D semantic bridging, cross-modal semantic guidance, and semantic-based attention mechanisms. Experimental validation on the BraTS 2020 dataset demonstrates that the proposed model achieves an overall Dice coefficient of 0.8567, representing a 4.8% improvement compared to traditional 3D U-Net, with a 7.3% Dice coefficient increase in the clinically important enhancing tumor (ET) region.
comment: 13 pages,6 figures
♻ ☆ Signal Recovery Using a Spiked Mixture Model
We introduce the spiked mixture model (SMM) to address the problem of estimating a set of signals from many randomly scaled and noisy observations. Subsequently, we design a novel expectation-maximization (EM) algorithm to recover all parameters of the SMM. Numerical experiments show that in low signal-to-noise ratio regimes, and for data types where the SMM is relevant, SMM surpasses the more traditional Gaussian mixture model (GMM) in terms of signal recovery performance. The broad relevance of the SMM and its corresponding EM recovery algorithm is demonstrated by applying the technique to different data types. The first case study is a biomedical research application, utilizing an imaging mass spectrometry dataset to explore the molecular content of a rat brain tissue section at micrometer scale. The second case study demonstrates SMM performance in a computer vision application, segmenting a hyperspectral imaging dataset into underlying patterns. While the measurement modalities differ substantially, in both case studies SMM is shown to recover signals that were missed by traditional methods such as k-means clustering and GMM.
♻ ☆ MMOne: Representing Multiple Modalities in One Scene ICCV 2025
Humans perceive the world through multimodal cues to understand and interact with the environment. Learning a scene representation for multiple modalities enhances comprehension of the physical world. However, modality conflicts, arising from inherent distinctions among different modalities, present two critical challenges: property disparity and granularity disparity. To address these challenges, we propose a general framework, MMOne, to represent multiple modalities in one scene, which can be readily extended to additional modalities. Specifically, a modality modeling module with a novel modality indicator is proposed to capture the unique properties of each modality. Additionally, we design a multimodal decomposition mechanism to separate multi-modal Gaussians into single-modal Gaussians based on modality differences. We address the essential distinctions among modalities by disentangling multimodal information into shared and modality-specific components, resulting in a more compact and efficient multimodal scene representation. Extensive experiments demonstrate that our method consistently enhances the representation capability for each modality and is scalable to additional modalities. The code is available at https://github.com/Neal2020GitHub/MMOne.
comment: Accepted to ICCV 2025
♻ ☆ Global urban visual perception varies across demographics and personalities
Understanding people's preferences is crucial for urban planning, yet current approaches often combine responses from multi-cultural populations, obscuring demographic differences and risking amplifying biases. We conducted a large-scale urban visual perception survey of streetscapes worldwide using street view imagery, examining how demographics -- including gender, age, income, education, race and ethnicity, and, for the first time, personality traits -- shape perceptions among 1,000 participants with balanced demographics from five countries and 45 nationalities. This dataset, Street Perception Evaluation Considering Socioeconomics (SPECS), reveals demographic- and personality-based differences across six traditional indicators (safe, lively, wealthy, beautiful, boring, depressing) and four new ones (live nearby, walk, cycle, green). Location-based sentiments further shape these preferences. Machine learning models trained on existing global datasets tend to overestimate positive indicators and underestimate negative ones compared to human responses, underscoring the need for local context. Our study aspires to rectify the myopic treatment of street perception, which rarely considers demographics or personality traits.
comment: Under review
♻ ☆ A column generation algorithm with dynamic constraint aggregation for minimum sum-of-squares clustering
The minimum sum-of-squares clustering problem (MSSC), also known as $k$-means clustering, refers to the problem of partitioning $n$ data points into $k$ clusters, with the objective of minimizing the total sum of squared Euclidean distances between each point and the center of its assigned cluster. We propose an efficient algorithm for solving large-scale MSSC instances, which combines column generation (CG) with dynamic constraint aggregation (DCA) to effectively reduce the number of constraints considered in the CG master problem. DCA was originally conceived to reduce degeneracy in set partitioning problems by utilizing an aggregated restricted master problem obtained from a partition of the set partitioning constraints into disjoint clusters. In this work, we explore the use of DCA within a CG algorithm for MSSC exact solution. Our method is fine-tuned by a series of ablation studies on DCA design choices, and is demonstrated to significantly outperform existing state-of-the-art exact approaches available in the literature.
♻ ☆ Latent Diffusion Model Based Denoising Receiver for 6G Semantic Communication: From Stochastic Differential Theory to Application
In this paper, a novel semantic communication framework empowered by generative artificial intelligence (GAI) is proposed, to enhance the robustness against both channel noise and transmission data distribution shifts. A theoretical foundation is established using stochastic differential equations (SDEs), from which a closed-form mapping between any signal-to-noise ratio (SNR) and the optimal denoising timestep is derived. Moreover, to address distribution mismatch, a mathematical scaling method is introduced to align received semantic features with the training distribution of the GAI. Built on this theoretical foundation, a latent diffusion model (LDM)-based semantic communication framework is proposed that combines a variational autoencoder for semantic features extraction, where a pretrained diffusion model is used for denoising. The proposed system is a training-free framework that supports zero-shot generalization, and achieves superior performance under low-SNR and out-of-distribution conditions, offering a scalable and robust solution for future 6G semantic communication systems. Experimental results demonstrate that the proposed semantic communication framework achieves state-of-the-art performance in both pixel-level accuracy and semantic perceptual quality, consistently outperforming baselines across a wide range of SNRs and data distributions without any fine-tuning or post-training.
♻ ☆ Site-Level Fine-Tuning with Progressive Layer Freezing: Towards Robust Prediction of Bronchopulmonary Dysplasia from Day-1 Chest Radiographs in Extremely Preterm Infants
Bronchopulmonary dysplasia (BPD) is a chronic lung disease affecting 35% of extremely low birth weight infants. Defined by oxygen dependence at 36 weeks postmenstrual age, it causes lifelong respiratory complications. However, preventive interventions carry severe risks, including neurodevelopmental impairment, ventilator-induced lung injury, and systemic complications. Therefore, early BPD prognosis and prediction of BPD outcome is crucial to avoid unnecessary toxicity in low risk infants. Admission radiographs of extremely preterm infants are routinely acquired within 24h of life and could serve as a non-invasive prognostic tool. In this work, we developed and investigated a deep learning approach using chest X-rays from 163 extremely low-birth-weight infants ($\leq$32 weeks gestation, 401-999g) obtained within 24 hours of birth. We fine-tuned a ResNet-50 pretrained specifically on adult chest radiographs, employing progressive layer freezing with discriminative learning rates to prevent overfitting and evaluated a CutMix augmentation and linear probing. For moderate/severe BPD outcome prediction, our best performing model with progressive freezing, linear probing and CutMix achieved an AUROC of 0.78 $\pm$ 0.10, balanced accuracy of 0.69 $\pm$ 0.10, and an F1-score of 0.67 $\pm$ 0.11. In-domain pre-training significantly outperformed ImageNet initialization (p = 0.031) which confirms domain-specific pretraining to be important for BPD outcome prediction. Routine IRDS grades showed limited prognostic value (AUROC 0.57 $\pm$ 0.11), confirming the need of learned markers. Our approach demonstrates that domain-specific pretraining enables accurate BPD prediction from routine day-1 radiographs. Through progressive freezing and linear probing, the method remains computationally feasible for site-level implementation and future federated learning deployments.
comment: S.G.-F., M.B., and A.E. contributed equally to this work and share first authorship. M.Z. and P.F. contributed equally to this work and share senior authorship
♻ ☆ Formalising causal inference as prediction on a target population ICML 2024
The standard approach to causal modelling especially in social and health sciences is the potential outcomes framework due to Neyman and Rubin. In this framework, observations are thought to be drawn from a distribution over variables of interest, and the goal is to identify parameters of this distribution. Even though the stated goal is often to inform decision making on some target population, there is no straightforward way to include these target populations in the framework. Instead of modelling the relationship between the observed sample and the target population, the inductive assumptions in this framework take the form of abstract sampling and independence assumptions. In this paper, we develop a version of this framework that construes causal inference as treatment-wise predictions for finite populations where all assumptions are testable in retrospect; this means that one can not only test predictions themselves (without any fundamental problem) but also investigate sources of error when they fail. Due to close connections to the original framework, established methods can still be be analysed under the new framework.
comment: Presented at the Humans, Algorithmic Decision-Making and Society Workshop at ICML 2024
♻ ☆ Dataset resulting from the user study on comprehensibility of explainable AI algorithms
This paper introduces a dataset that is the result of a user study on the comprehensibility of explainable artificial intelligence (XAI) algorithms. The study participants were recruited from 149 candidates to form three groups representing experts in the domain of mycology (DE), students with a data science and visualization background (IT) and students from social sciences and humanities (SSH). The main part of the dataset contains 39 transcripts of interviews during which participants were asked to complete a series of tasks and questions related to the interpretation of explanations of decisions of a machine learning model trained to distinguish between edible and inedible mushrooms. The transcripts were complemented with additional data that includes visualizations of explanations presented to the user, results from thematic analysis, recommendations of improvements of explanations provided by the participants, and the initial survey results that allow to determine the domain knowledge of the participant and data analysis literacy. The transcripts were manually tagged to allow for automatic matching between the text and other data related to particular fragments. In the advent of the area of rapid development of XAI techniques, the need for a multidisciplinary qualitative evaluation of explainability is one of the emerging topics in the community. Our dataset allows not only to reproduce the study we conducted, but also to open a wide range of possibilities for the analysis of the material we gathered.
♻ ☆ Task-Specific Generative Dataset Distillation with Difficulty-Guided Sampling ICCV 2025
To alleviate the reliance of deep neural networks on large-scale datasets, dataset distillation aims to generate compact, high-quality synthetic datasets that can achieve comparable performance to the original dataset. The integration of generative models has significantly advanced this field. However, existing approaches primarily focus on aligning the distilled dataset with the original one, often overlooking task-specific information that can be critical for optimal downstream performance. In this paper, focusing on the downstream task of classification, we propose a task-specific sampling strategy for generative dataset distillation that incorporates the concept of difficulty to consider the requirements of the target task better. The final dataset is sampled from a larger image pool with a sampling distribution obtained by matching the difficulty distribution of the original dataset. A logarithmic transformation is applied as a pre-processing step to correct for distributional bias. The results of extensive experiments demonstrate the effectiveness of our method and suggest its potential for enhancing performance on other downstream tasks. The code is available at https://github.com/SumomoTaku/DiffGuideSamp.
comment: Accepted by The ICCV 2025 Workshop on Curated Data for Efficient Learning
♻ ☆ We should avoid the assumption of data-generating probability distributions in social settings ICML 2024
Machine Learning research, including work promoting fair or equitable algorithms, heavily relies on the concept of a data-generating probability distribution. The standard presumption is that since data points are 'sampled from' such a distribution, one can learn from observed data about this distribution and, thus, predict future data points which are also drawn from it. We argue, however, that such true probability distributions do not exist and should not be dealt with uncritically. We show that alternative frameworks focusing directly on relevant populations rather than abstract distributions are available and leave classical learning theory almost unchanged. Furthermore, we argue that the assumption of true probabilities or data-generating distributions can be misleading and obscure both the choices made and the goals pursued in machine learning practice. Based on these considerations, this position paper argues that, at least in social settings, machine learning work should avoid assuming data-generating probability distributions.
comment: Presented at the Humans, Algorithmic Decision-Making and Society Workshop at ICML 2024
♻ ☆ ZClassifier: Temperature Tuning and Manifold Approximation via KL Divergence on Logit Space
We introduce a novel classification framework, ZClassifier, that replaces conventional deterministic logits with diagonal Gaussian-distributed logits. Our method simultaneously addresses temperature scaling and manifold approximation by minimizing the Kullback-Leibler (KL) divergence between the predicted Gaussian distributions and a unit isotropic Gaussian. This unifies uncertainty calibration and latent control in a principled probabilistic manner, enabling a natural interpretation of class confidence and geometric consistency. Experiments on CIFAR-10 show that ZClassifier improves over softmax classifiers in robustness, calibration, and latent separation.
♻ ☆ Holistix: A Dataset for Holistic Wellness Dimensions Analysis in Mental Health Narratives
We introduce a dataset for classifying wellness dimensions in social media user posts, covering six key aspects: physical, emotional, social, intellectual, spiritual, and vocational. The dataset is designed to capture these dimensions in user-generated content, with a comprehensive annotation framework developed under the guidance of domain experts. This framework allows for the classification of text spans into the appropriate wellness categories. We evaluate both traditional machine learning models and advanced transformer-based models for this multi-class classification task, with performance assessed using precision, recall, and F1-score, averaged over 10-fold cross-validation. Post-hoc explanations are applied to ensure the transparency and interpretability of model decisions. The proposed dataset contributes to region-specific wellness assessments in social media and paves the way for personalized well-being evaluations and early intervention strategies in mental health. We adhere to ethical considerations for constructing and releasing our experiments and dataset publicly on Github.
comment: 7 Pages
♻ ☆ Quantum Long Short-Term Memory for Drug Discovery
Quantum computing combined with machine learning (ML) is a highly promising research area, with numerous studies demonstrating that quantum machine learning (QML) is expected to solve scientific problems more effectively than classical ML. In this work, we present Quantum Long Short-Term Memory (QLSTM), a QML architecture, and demonstrate its effectiveness in drug discovery. We evaluate QLSTM on five benchmark datasets (BBBP, BACE, SIDER, BCAP37, T-47D), and observe consistent performance gains over classical LSTM, with ROC-AUC improvements ranging from 3% to over 6%. Furthermore, QLSTM exhibits improved predictive accuracy as the number of qubits increases, and faster convergence than classical LSTM under the same training conditions. Notably, QLSTM maintains strong robustness against quantum computer noise, outperforming noise-free classical LSTM in certain settings. These findings highlight the potential of QLSTM as a scalable and noise-resilient model for scientific applications, particularly as quantum hardware continues to advance in qubit capacity and fidelity.
♻ ☆ Deep Q-Learning with Gradient Target Tracking
This paper introduces Q-learning with gradient target tracking, a novel reinforcement learning framework that provides a learned continuous target update mechanism as an alternative to the conventional hard update paradigm. In the standard deep Q-network (DQN), the target network is a copy of the online network's weights, held fixed for a number of iterations before being periodically replaced via a hard update. While this stabilizes training by providing consistent targets, it introduces a new challenge: the hard update period must be carefully tuned to achieve optimal performance. To address this issue, we propose two gradient-based target update methods: DQN with asymmetric gradient target tracking (AGT2-DQN) and DQN with symmetric gradient target tracking (SGT2-DQN). These methods replace the conventional hard target updates with continuous and structured updates using gradient descent, which effectively eliminates the need for manual tuning. We provide a theoretical analysis proving the convergence of these methods in tabular settings. Additionally, empirical evaluations demonstrate their advantages over standard DQN baselines, which suggest that gradient-based target updates can serve as an effective alternative to conventional target update mechanisms in Q-learning.
♻ ☆ Physics-Informed Linear Model (PILM): Analytical Representations and Application to Crustal Strain Rate Estimation
Many physical systems are described by partial differential equations (PDEs), and solving these equations and estimating their coefficients or boundary conditions (BCs) from observational data play a crucial role in understanding the associated phenomena. Recently, a machine learning approach known as physics-informed neural network, which solves PDEs using neural networks by minimizing the sum of residuals from the PDEs, BCs, and data, has gained significant attention in the scientific community. In this study, we investigate a physics-informed linear model (PILM) that uses linear combinations of basis functions to represent solutions, thereby enabling an analytical representation of optimal solutions. The PILM was formulated and verified for illustrative forward and inverse problems including cases with uncertain BCs. Furthermore, the PILM was applied to estimate crustal strain rates using geodetic data. Specifically, physical regularization that enforces elastic equilibrium on the velocity fields was compared with mathematical regularization that imposes smoothness constraints. From a Bayesian perspective, mathematical regularization exhibited superior performance. The PILM provides an analytically solvable framework applicable to linear forward and inverse problems, underdetermined systems, and physical regularization.
♻ ☆ ReCode: Updating Code API Knowledge with Reinforcement Learning
Large Language Models (LLMs) exhibit remarkable code generation capabilities but falter when adapting to frequent updates in external library APIs. This critical limitation, stemming from reliance on outdated API knowledge from their training data, even with access to current documentation, impedes reliable code generation in dynamic environments. To tackle this issue, we propose ReCode (rule-based Reinforcement learning for Code Update), a novel framework that mimics human programmer adaptation to API changes. Specifically, we construct a dataset of approximately 2,000 data entries to train the LLMs to perform version migration based on updated information. Then, we introduce a modified string similarity metric for code evaluation as the reward for reinforcement learning. Our experiments demonstrate that ReCode substantially boosts LLMs' code generation performance in dynamic API scenarios, especially on the unseen CodeUpdateArena task. Crucially, compared to supervised fine-tuning, ReCode has less impact on LLMs' general code generation abilities. We apply ReCode on various LLMs and reinforcement learning algorithms (GRPO and DAPO), all achieving consistent improvements. Notably, after training, Qwen2.5-Coder-7B outperforms that of the 32B parameter code instruction-tuned model and the reasoning model with the same architecture. Code is available at https://github.com/zjunlp/ReCode.
comment: Work in progress
♻ ☆ MPO: An Efficient Post-Processing Framework for Mixing Diverse Preference Alignment ICML 2025
Reinforcement Learning from Human Feedback (RLHF) has shown promise in aligning large language models (LLMs). Yet its reliance on a singular reward model often overlooks the diversity of human preferences. Recent approaches address this limitation by leveraging multi-dimensional feedback to fine-tune corresponding reward models and train LLMs using reinforcement learning. However, the process is costly and unstable, especially given the competing and heterogeneous nature of human preferences. In this paper, we propose Mixing Preference Optimization (MPO), a post-processing framework for aggregating single-objective policies as an alternative to both multi-objective RLHF (MORLHF) and MaxMin-RLHF. MPO avoids alignment from scratch. Instead, it log-linearly combines existing policies into a unified one with the weight of each policy computed via a batch stochastic mirror descent. Empirical results demonstrate that MPO achieves balanced performance across diverse preferences, outperforming or matching existing models with significantly reduced computational costs.
comment: ICML 2025
♻ ☆ Demystifying MuZero Planning: Interpreting the Learned Model
MuZero has achieved superhuman performance in various games by using a dynamics network to predict the environment dynamics for planning, without relying on simulators. However, the latent states learned by the dynamics network make its planning process opaque. This paper aims to demystify MuZero's model by interpreting the learned latent states. We incorporate observation reconstruction and state consistency into MuZero training and conduct an in-depth analysis to evaluate latent states across two board games: 9x9 Go and Gomoku, and three Atari games: Breakout, Ms. Pacman, and Pong. Our findings reveal that while the dynamics network becomes less accurate over longer simulations, MuZero still performs effectively by using planning to correct errors. Our experiments also show that the dynamics network learns better latent states in board games than in Atari games. These insights contribute to a better understanding of MuZero and offer directions for future research to improve the performance, robustness, and interpretability of the MuZero algorithm. The code and data are available at https://rlg.iis.sinica.edu.tw/papers/demystifying-muzero-planning.
comment: Accepted by IEEE Transactions on Artificial Intelligence
♻ ☆ Golden Noise for Diffusion Models: A Learning Framework
Text-to-image diffusion model is a popular paradigm that synthesizes personalized images by providing a text prompt and a random Gaussian noise. While people observe that some noises are ``golden noises'' that can achieve better text-image alignment and higher human preference than others, we still lack a machine learning framework to obtain those golden noises. To learn golden noises for diffusion sampling, we mainly make three contributions in this paper. First, we identify a new concept termed the \textit{noise prompt}, which aims at turning a random Gaussian noise into a golden noise by adding a small desirable perturbation derived from the text prompt. Following the concept, we first formulate the \textit{noise prompt learning} framework that systematically learns ``prompted'' golden noise associated with a text prompt for diffusion models. Second, we design a noise prompt data collection pipeline and collect a large-scale \textit{noise prompt dataset}~(NPD) that contains 100k pairs of random noises and golden noises with the associated text prompts. With the prepared NPD as the training dataset, we trained a small \textit{noise prompt network}~(NPNet) that can directly learn to transform a random noise into a golden noise. The learned golden noise perturbation can be considered as a kind of prompt for noise, as it is rich in semantic information and tailored to the given text prompt. Third, our extensive experiments demonstrate the impressive effectiveness and generalization of NPNet on improving the quality of synthesized images across various diffusion models, including SDXL, DreamShaper-xl-v2-turbo, and Hunyuan-DiT. Moreover, NPNet is a small and efficient controller that acts as a plug-and-play module with very limited additional inference and computational costs, as it just provides a golden noise instead of a random noise without accessing the original pipeline.
♻ ☆ TBDetector:Transformer-Based Detector for Advanced Persistent Threats with Provenance Graph
APT detection is difficult to detect due to the long-term latency, covert and slow multistage attack patterns of Advanced Persistent Threat (APT). To tackle these issues, we propose TBDetector, a transformer-based advanced persistent threat detection method for APT attack detection. Considering that provenance graphs provide rich historical information and have the powerful attacks historic correlation ability to identify anomalous activities, TBDetector employs provenance analysis for APT detection, which summarizes long-running system execution with space efficiency and utilizes transformer with self-attention based encoder-decoder to extract long-term contextual features of system states to detect slow-acting attacks. Furthermore, we further introduce anomaly scores to investigate the anomaly of different system states, where each state is calculated with an anomaly score corresponding to its similarity score and isolation score. To evaluate the effectiveness of the proposed method, we have conducted experiments on five public datasets, i.e., streamspot, cadets, shellshock, clearscope, and wget_baseline. Experimental results and comparisons with state-of-the-art methods have exhibited better performance of our proposed method.
comment: 10 pages, 7 figures
♻ ☆ Faster and Space Efficient Indexing for Locality Sensitive Hashing
This work suggests faster and space-efficient index construction algorithms for LSH for Euclidean distance (\textit{a.k.a.}~\ELSH) and cosine similarity (\textit{a.k.a.}~\SRP). The index construction step of these LSHs relies on grouping data points into several bins of hash tables based on their hashcode. To generate an $m$-dimensional hashcode of the $d$-dimensional data point, these LSHs first project the data point onto a $d$-dimensional random Gaussian vector and then discretise the resulting inner product. The time and space complexity of both \ELSH~and \SRP~for computing an $m$-sized hashcode of a $d$-dimensional vector is $O(md)$, which becomes impractical for large values of $m$ and $d$. To overcome this problem, we propose two alternative LSH hashcode generation algorithms, both for Euclidean distance and cosine similarity, namely, \CSELSH, \HCSELSH~and \CSSRP, \HCSSRP, respectively. \CSELSH~and \CSSRP~are based on count sketch \cite{count_sketch} and \HCSELSH~and \HCSSRP~utilize higher-order count sketch \cite{shi2019higher}. These proposals significantly reduce the hashcode computation time from $O(md)$ to $O(d)$. Additionally, both \CSELSH~and \CSSRP~reduce the space complexity from $O(md)$ to $O(d)$; ~and \HCSELSH, \HCSSRP~ reduce the space complexity from $O(md)$ to $O(N \sqrt[N]{d})$ respectively, where $N\geq 1$ denotes the size of the input/reshaped tensor. Our proposals are backed by strong mathematical guarantees, and we validate their performance through simulations on various real-world datasets.
♻ ☆ A Comprehensive Survey of Synthetic Tabular Data Generation
Tabular data is one of the most prevalent and important data formats in real-world applications such as healthcare, finance, and education. However, its effective use in machine learning is often constrained by data scarcity, privacy concerns, and class imbalance. Synthetic tabular data generation has emerged as a powerful solution, leveraging generative models to learn underlying data distributions and produce realistic, privacy-preserving samples. Although this area has seen growing attention, most existing surveys focus narrowly on specific methods (e.g., GANs or privacy-enhancing techniques), lacking a unified and comprehensive view that integrates recent advances such as diffusion models and large language models (LLMs). In this survey, we present a structured and in-depth review of synthetic tabular data generation methods. Specifically, the survey is organized into three core components: (1) Background, which covers the overall generation pipeline, including problem definitions, synthetic tabular data generation methods, post processing, and evaluation; (2) Generation Methods, where we categorize existing approaches into traditional generation methods, diffusion model methods, and LLM-based methods, and compare them in terms of architecture, generation quality, and applicability; and (3) Applications and Challenges, which summarizes practical use cases, highlights common datasets, and discusses open challenges such as heterogeneity, data fidelity, and privacy protection. This survey aims to provide researchers and practitioners with a holistic understanding of the field and to highlight key directions for future work in synthetic tabular data generation.
♻ ☆ Learning Universal Human Mobility Patterns with a Foundation Model for Cross-domain Data Fusion
Human mobility modeling is critical for urban planning and transportation management, yet existing approaches often lack the integration capabilities needed to handle diverse data sources. We present a foundation model framework for universal human mobility patterns that leverages cross-domain data fusion and large language models to address these limitations. Our approach integrates multi-modal data of distinct nature and spatio-temporal resolution, including geographical, mobility, socio-demographic, and traffic information, to construct a privacy-preserving and semantically enriched human travel trajectory dataset. Our framework demonstrates adaptability through domain transfer techniques that ensure transferability across diverse urban contexts, as evidenced in case studies of Los Angeles (LA) and Egypt. The framework employs LLMs for semantic enrichment of trajectory data, enabling comprehensive understanding of mobility patterns. Quantitative evaluation shows that our generated synthetic dataset accurately reproduces mobility patterns observed in empirical data. The practical utility of this foundation model approach is demonstrated through large-scale traffic simulations for LA County, where results align well with observed traffic data. On California's I-405 corridor, the simulation yields a Mean Absolute Percentage Error of 5.85% for traffic volume and 4.36% for speed compared to Caltrans PeMS observations, illustrating the framework's potential for intelligent transportation systems and urban mobility applications.
♻ ☆ How does Labeling Error Impact Contrastive Learning? A Perspective from Data Dimensionality Reduction ICML2025
In recent years, contrastive learning has achieved state-of-the-art performance in the territory of self-supervised representation learning. Many previous works have attempted to provide the theoretical understanding underlying the success of contrastive learning. Almost all of them rely on a default assumption, i.e., the label consistency assumption, which may not hold in practice (the probability of failure is called labeling error) due to the strength and randomness of common augmentation strategies, such as random resized crop (RRC). This paper investigates the theoretical impact of labeling error on the downstream classification performance of contrastive learning. We first reveal several significant negative impacts of labeling error on downstream classification risk. To mitigate these impacts, data dimensionality reduction method (e.g., singular value decomposition, SVD) is applied on original data to reduce false positive samples, and establish both theoretical and empirical evaluations. Moreover, it is also found that SVD acts as a double-edged sword, which may lead to the deterioration of downstream classification accuracy due to the reduced connectivity of the augmentation graph. Based on the above observations, we give the augmentation suggestion that we should use some moderate embedding dimension (such as $512, 1024$ in our experiments), data inflation, weak augmentation, and SVD to ensure large graph connectivity and small labeling error to improve model performance.
comment: Published as ICML2025 poster. The arXiv version is a modified version
♻ ☆ Unifying Explainable Anomaly Detection and Root Cause Analysis in Dynamical Systems AAAI-25
Dynamical systems, prevalent in various scientific and engineering domains, are susceptible to anomalies that can significantly impact their performance and reliability. This paper addresses the critical challenges of anomaly detection, root cause localization, and anomaly type classification in dynamical systems governed by ordinary differential equations (ODEs). We define two categories of anomalies: cyber anomalies, which propagate through interconnected variables, and measurement anomalies, which remain localized to individual variables. To address these challenges, we propose the Interpretable Causality Ordinary Differential Equation (ICODE) Networks, a model-intrinsic explainable learning framework. ICODE leverages Neural ODEs for anomaly detection while employing causality inference through an explanation channel to perform root cause analysis (RCA), elucidating why specific time periods are flagged as anomalous. ICODE is designed to simultaneously perform anomaly detection, RCA, and anomaly type classification within a single, interpretable framework. Our approach is grounded in the hypothesis that anomalies alter the underlying ODEs of the system, manifesting as changes in causal relationships between variables. We provide a theoretical analysis of how perturbations in learned model parameters can be utilized to identify anomalies and their root causes in time series data. Comprehensive experimental evaluations demonstrate the efficacy of ICODE across various dynamical systems, showcasing its ability to accurately detect anomalies, classify their types, and pinpoint their origins.
comment: Accepted by the AAAI-25 Workshop on Artificial Intelligence for Cyber Security (AICS)
♻ ☆ BEARCUBS: A benchmark for computer-using web agents
Modern web agents possess computer use abilities that allow them to interact with webpages by sending commands to a virtual keyboard and mouse. While such agents have considerable potential to assist human users with complex tasks, evaluating their capabilities in real-world settings poses a major challenge. To this end, we introduce BEARCUBS, a "small but mighty" benchmark of 111 information-seeking questions designed to evaluate a web agent's ability to search, browse, and identify factual information from the web. Unlike prior web agent benchmarks, solving BEARCUBS requires (1) accessing live web content rather than synthetic or simulated pages, which captures the unpredictability of real-world web interactions; and (2) performing a broad range of multimodal interactions (e.g., video understanding, 3D navigation) that cannot be bypassed via text-based workarounds. Each question in BEARCUBS has a corresponding short, unambiguous answer and a human-validated browsing trajectory, allowing for transparent evaluation of agent performance and strategies. A human study confirms that BEARCUBS questions are solvable but non-trivial (84.7% human accuracy), revealing domain knowledge gaps and overlooked details as common failure points. By contrast, state-of-the-art computer-using agents underperform, with the best-scoring system (OpenAI's Operator) reaching only 23.4% accuracy. These results highlight critical areas for improvement, including reliable source selection and more powerful multimodal capabilities. To facilitate future research, BEARCUBS will be updated periodically to replace invalid or contaminated questions, keeping the benchmark fresh for future generations of web agents.
comment: 16 pages
♻ ☆ Rethinking Inductive Bias in Geographically Neural Network Weighted Regression
Inductive bias is a key factor in spatial regression models, determining how well a model can learn from limited data and capture spatial patterns. This work revisits the inductive biases in Geographically Neural Network Weighted Regression (GNNWR) and identifies limitations in current approaches for modeling spatial non-stationarity. While GNNWR extends traditional Geographically Weighted Regression by using neural networks to learn spatial weighting functions, existing implementations are often restricted by fixed distance-based schemes and limited inductive bias. We propose to generalize GNNWR by incorporating concepts from convolutional neural networks, recurrent neural networks, and transformers, introducing local receptive fields, sequential context, and self-attention into spatial regression. Through extensive benchmarking on synthetic spatial datasets with varying heterogeneity, noise, and sample sizes, we show that GNNWR outperforms classic methods in capturing nonlinear and complex spatial relationships. Our results also reveal that model performance depends strongly on data characteristics, with local models excelling in highly heterogeneous or small-sample scenarios, and global models performing better with larger, more homogeneous data. These findings highlight the importance of inductive bias in spatial modeling and suggest future directions, including learnable spatial weighting functions, hybrid neural architectures, and improved interpretability for models handling non-stationary spatial data.
♻ ☆ Multi-View Node Pruning for Accurate Graph Representation
Graph pooling, which compresses a whole graph into a smaller coarsened graph, is an essential component of graph representation learning. To efficiently compress a given graph, graph pooling methods often drop their nodes with attention-based scoring with the task loss. However, this often results in simply removing nodes with lower degrees without consideration of their feature-level relevance to the given task. To fix this problem, we propose a Multi-View Pruning(MVP), a graph pruning method based on a multi-view framework and reconstruction loss. Given a graph, MVP first constructs multiple graphs for different views either by utilizing the predefined modalities or by randomly partitioning the input features, to consider the importance of each node in diverse perspectives. Then, it learns the score for each node by considering both the reconstruction and the task loss. MVP can be incorporated with any hierarchical pooling framework to score the nodes. We validate MVP on multiple benchmark datasets by coupling it with two graph pooling methods, and show that it significantly improves the performance of the base graph pooling method, outperforming all baselines. Further analysis shows that both the encoding of multiple views and the consideration of reconstruction loss are the key to the success of MVP, and that it indeed identifies nodes that are less important according to domain knowledge.
comment: Jiseong Park and Hanjin Kim are co-first author for this work
♻ ☆ Scaling Trends for Data Poisoning in LLMs
LLMs produce harmful and undesirable behavior when trained on datasets containing even a small fraction of poisoned data. We demonstrate that GPT models remain vulnerable to fine-tuning on poisoned data, even when safeguarded by moderation systems. Given the persistence of data poisoning vulnerabilities in today's most capable models, this paper investigates whether these risks increase with model scaling. We evaluate three threat models -- malicious fine-tuning, imperfect data curation, and intentional data contamination -- across 24 frontier LLMs ranging from 1.5 to 72 billion parameters. Our experiments reveal that larger LLMs are significantly more susceptible to data poisoning, learning harmful behaviors from even minimal exposure to harmful data more quickly than smaller models. These findings underscore the need for leading AI companies to thoroughly red team fine-tuning APIs before public release and to develop more robust safeguards against data poisoning, particularly as models continue to scale in size and capability.
comment: This arXiv version of the paper originally included an initial investigation of jailbreak-tuning, which can produce 60+ percentage point increases in vulnerability elicitation compared with standard data poisoning. Jailbreak-tuning has now been separated into a full independent paper, which can be found at arXiv:2507.11630
Multimedia 11
☆ Bidirectional Age of Incorrect Information: A Performance Metric for Status Updates in Virtual Dynamic Environments
Virtual dynamic environments (VDEs) such as the Metaverse and digital twins (DTs) require proper representation of the interacting entities to map their characteristics within the simulated or augmented space. Keeping these representations accurate and up-to-date is crucial for seamless interaction and system reliability. In this paper, we propose bidirectional age of incorrect information (BAoII) to address this aspect. BAoII quantifies the time-dependent penalty paid by an entity in a VDE due to incorrect or outdated knowledge about itself and the overall dynamically changing space. This extends the concept of age of incorrect information for a bidirectional information exchange, capturing that a VDE requires mutual awareness of the entity's own representation, measured in the virtual space, and what the other entities share about their representations. Using a continuous-time Markov chain model, we derive a closed-form expression for long-term BAoII and identify a transmission cost threshold for optimal update strategies. We describe a trade-off between communication cost and information freshness and validate our model through numerical simulations, demonstrating the impact of BAoII on evaluating system performance and highlighting its relevance for real-time collaboration in the Metaverse and DTs.
comment: 8 pages, 8 figures, 1 table, Proc. IEEE Metacom
☆ Automating Steering for Safe Multimodal Large Language Models
Recent progress in Multimodal Large Language Models (MLLMs) has unlocked powerful cross-modal reasoning abilities, but also raised new safety concerns, particularly when faced with adversarial multimodal inputs. To improve the safety of MLLMs during inference, we introduce a modular and adaptive inference-time intervention technology, AutoSteer, without requiring any fine-tuning of the underlying model. AutoSteer incorporates three core components: (1) a novel Safety Awareness Score (SAS) that automatically identifies the most safety-relevant distinctions among the model's internal layers; (2) an adaptive safety prober trained to estimate the likelihood of toxic outputs from intermediate representations; and (3) a lightweight Refusal Head that selectively intervenes to modulate generation when safety risks are detected. Experiments on LLaVA-OV and Chameleon across diverse safety-critical benchmarks demonstrate that AutoSteer significantly reduces the Attack Success Rate (ASR) for textual, visual, and cross-modal threats, while maintaining general abilities. These findings position AutoSteer as a practical, interpretable, and effective framework for safer deployment of multimodal AI systems.
comment: Working in progress. 22 pages (8+ for main); 25 figures; 1 table
☆ Predictability-Aware Motion Prediction for Edge XR via High-Order Error-State Kalman Filtering
As 6G networks are developed and defined, offloading of XR applications is emerging as one of the strong new use cases. The reduced 6G latency coupled with edge processing infrastructure will for the first time provide a realistic offloading scenario in cellular networks where several computationally intensive functions, including rendering, can migrate from the user device and into the network. A key advantage of doing so is the lowering of the battery needs in the user devices and the possibility to design new devices with smaller form factors.
☆ UniSLU: Unified Spoken Language Understanding from Heterogeneous Cross-Task Datasets
Spoken Language Understanding (SLU) plays a crucial role in speech-centric multimedia applications, enabling machines to comprehend spoken language in scenarios such as meetings, interviews, and customer service interactions. SLU encompasses multiple tasks, including Automatic Speech Recognition (ASR), spoken Named Entity Recognition (NER), and spoken Sentiment Analysis (SA). However, existing methods often rely on separate model architectures for individual tasks such as spoken NER and SA, which increases system complexity, limits cross-task interaction, and fails to fully exploit heterogeneous datasets available across tasks. To address these limitations, we propose UniSLU, a unified framework that jointly models multiple SLU tasks within a single architecture. Specifically, we propose a unified representation for diverse SLU tasks, enabling full utilization of heterogeneous datasets across multiple tasks. Built upon this representation, we propose a unified generative method that jointly models ASR, spoken NER, and SA tasks, enhancing task interactions and enabling seamless integration with large language models to harness their powerful generative capabilities. Extensive experiments on public SLU datasets demonstrate the effectiveness of our approach, achieving superior SLU performance compared to several benchmark methods, making it well-suited for real-world speech-based multimedia scenarios. We will release all code and models at github to facilitate future research.
comment: 13 pages, 3 figures
☆ Enkidu: Universal Frequential Perturbation for Real-Time Audio Privacy Protection against Voice Deepfakes ACM MM 2025
The rapid advancement of voice deepfake technologies has raised serious concerns about user audio privacy, as attackers increasingly exploit publicly available voice data to generate convincing fake audio for malicious purposes such as identity theft, financial fraud, and misinformation campaigns. While existing defense methods offer partial protection, they face critical limitations, including weak adaptability to unseen user data, poor scalability to long audio, rigid reliance on white-box knowledge, and high computational and temporal costs during the encryption process. To address these challenges and defend against personalized voice deepfake threats, we propose Enkidu, a novel user-oriented privacy-preserving framework that leverages universal frequential perturbations generated through black-box knowledge and few-shot training on a small amount of user data. These highly malleable frequency-domain noise patches enable real-time, lightweight protection with strong generalization across variable-length audio and robust resistance to voice deepfake attacks, all while preserving perceptual quality and speech intelligibility. Notably, Enkidu achieves over 50 to 200 times processing memory efficiency (as low as 0.004 gigabytes) and 3 to 7000 times runtime efficiency (real-time coefficient as low as 0.004) compared to six state-of-the-art countermeasures. Extensive experiments across six mainstream text-to-speech models and five cutting-edge automated speaker verification models demonstrate the effectiveness, transferability, and practicality of Enkidu in defending against both vanilla and adaptive voice deepfake attacks.
comment: Accepted by ACM MM 2025
☆ Cross-Modal Watermarking for Authentic Audio Recovery and Tamper Localization in Synthesized Audiovisual Forgeries
Recent advances in voice cloning and lip synchronization models have enabled Synthesized Audiovisual Forgeries (SAVFs), where both audio and visuals are manipulated to mimic a target speaker. This significantly increases the risk of misinformation by making fake content seem real. To address this issue, existing methods detect or localize manipulations but cannot recover the authentic audio that conveys the semantic content of the message. This limitation reduces their effectiveness in combating audiovisual misinformation. In this work, we introduce the task of Authentic Audio Recovery (AAR) and Tamper Localization in Audio (TLA) from SAVFs and propose a cross-modal watermarking framework to embed authentic audio into visuals before manipulation. This enables AAR, TLA, and a robust defense against misinformation. Extensive experiments demonstrate the strong performance of our method in AAR and TLA against various manipulations, including voice cloning and lip synchronization.
comment: 5 pages, 2 figures, Interspeech 2025
♻ ☆ KEN: Knowledge Augmentation and Emotion Guidance Network for Multimodal Fake News Detection ACM MM 2025
In recent years, the rampant spread of misinformation on social media has made accurate detection of multimodal fake news a critical research focus. However, previous research has not adequately understood the semantics of images, and models struggle to discern news authenticity with limited textual information. Meanwhile, treating all emotional types of news uniformly without tailored approaches further leads to performance degradation. Therefore, we propose a novel Knowledge Augmentation and Emotion Guidance Network (KEN). On the one hand, we effectively leverage LVLM's powerful semantic understanding and extensive world knowledge. For images, the generated captions provide a comprehensive understanding of image content and scenes, while for text, the retrieved evidence helps break the information silos caused by the closed and limited text and context. On the other hand, we consider inter-class differences between different emotional types of news through balanced learning, achieving fine-grained modeling of the relationship between emotional types and authenticity. Extensive experiments on two real-world datasets demonstrate the superiority of our KEN.
comment: Accepted by ACM MM 2025
♻ ☆ Benchmarking Sub-Genre Classification For Mainstage Dance Music SP
Music classification, a cornerstone of music information retrieval, supports a wide array of applications. To address the lack of comprehensive datasets and effective methods for sub-genre classification in mainstage dance music, we introduce a novel benchmark featuring a new dataset and baseline. Our dataset expands the scope of sub-genres to reflect the diversity of recent mainstage live sets performed by leading DJs at global music festivals, capturing the vibrant and rapidly evolving electronic dance music (EDM) scene that engages millions of fans worldwide. We employ a continuous soft labeling approach to accommodate tracks blending multiple sub-genres, preserving their inherent complexity. Experiments demonstrate that even state-of-the-art multimodal large language models (MLLMs) struggle with this task, while our specialized baseline models achieve high accuracy. This benchmark supports applications such as music recommendation, DJ set curation, and interactive multimedia systems, with video demos provided. Our code and data are all open-sourced at https://github.com/Gariscat/housex-v2.git}{https://github.com/Gariscat/housex-v2.git.
comment: WASPAA 2025
♻ ☆ Can Large Language Models Predict Audio Effects Parameters from Natural Language? SP
In music production, manipulating audio effects (Fx) parameters through natural language has the potential to reduce technical barriers for non-experts. We present LLM2Fx, a framework leveraging Large Language Models (LLMs) to predict Fx parameters directly from textual descriptions without requiring task-specific training or fine-tuning. Our approach address the text-to-effect parameter prediction (Text2Fx) task by mapping natural language descriptions to the corresponding Fx parameters for equalization and reverberation. We demonstrate that LLMs can generate Fx parameters in a zero-shot manner that elucidates the relationship between timbre semantics and audio effects in music production. To enhance performance, we introduce three types of in-context examples: audio Digital Signal Processing (DSP) features, DSP function code, and few-shot examples. Our results demonstrate that LLM-based Fx parameter generation outperforms previous optimization approaches, offering competitive performance in translating natural language descriptions to appropriate Fx settings. Furthermore, LLMs can serve as text-driven interfaces for audio production, paving the way for more intuitive and accessible music production tools.
comment: Accepted for publication at The IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA)
♻ ☆ Music Grounding by Short Video ICCV2025
Adding proper background music helps complete a short video to be shared. Previous work tackles the task by video-to-music retrieval (V2MR), aiming to find the most suitable music track from a collection to match the content of a given query video. In practice, however, music tracks are typically much longer than the query video, necessitating (manual) trimming of the retrieved music to a shorter segment that matches the video duration. In order to bridge the gap between the practical need for music moment localization and V2MR, we propose a new task termed Music Grounding by Short Video (MGSV). To tackle the new task, we introduce a new benchmark, MGSV-EC, which comprises a diverse set of 53k short videos associated with 35k different music moments from 4k unique music tracks. Furthermore, we develop a new baseline method, MaDe, which performs both video-to-music matching and music moment detection within a unified end-to-end deep network. Extensive experiments on MGSV-EC not only highlight the challenging nature of MGSV but also set MaDe as a strong baseline.
comment: Accepted to ICCV2025
♻ ☆ Fine-grained Image Retrieval via Dual-Vision Adaptation
Fine-Grained Image Retrieval~(FGIR) faces challenges in learning discriminative visual representations to retrieve images with similar fine-grained features. Current leading FGIR solutions typically follow two regimes: enforce pairwise similarity constraints in the semantic embedding space, or incorporate a localization sub-network to fine-tune the entire model. However, such two regimes tend to overfit the training data while forgetting the knowledge gained from large-scale pre-training, thus reducing their generalization ability. In this paper, we propose a Dual-Vision Adaptation (DVA) approach for FGIR, which guides the frozen pre-trained model to perform FGIR through collaborative sample and feature adaptation. Specifically, we design Object-Perceptual Adaptation, which modifies input samples to help the pre-trained model perceive critical objects and elements within objects that are helpful for category prediction. Meanwhile, we propose In-Context Adaptation, which introduces a small set of parameters for feature adaptation without modifying the pre-trained parameters. This makes the FGIR task using these adjusted features closer to the task solved during the pre-training. Additionally, to balance retrieval efficiency and performance, we propose Discrimination Perception Transfer to transfer the discriminative knowledge in the object-perceptual adaptation to the image encoder using the knowledge distillation mechanism. Extensive experiments show that DVA has fewer learnable parameters and performs well on three in-distribution and three out-of-distribution fine-grained datasets.
Computation and Language 99
☆ Language Models Improve When Pretraining Data Matches Target Tasks
Every data selection method inherently has a target. In practice, these targets often emerge implicitly through benchmark-driven iteration: researchers develop selection strategies, train models, measure benchmark performance, then refine accordingly. This raises a natural question: what happens when we make this optimization explicit? To explore this, we propose benchmark-targeted ranking (BETR), a simple method that selects pretraining documents based on similarity to benchmark training examples. BETR embeds benchmark examples and a sample of pretraining documents in a shared space, scores this sample by similarity to benchmarks, then trains a lightweight classifier to predict these scores for the full corpus. We compare data selection methods by training over 500 models spanning $10^{19}$ to $10^{22}$ FLOPs and fitting scaling laws to them. From this, we find that simply aligning pretraining data to evaluation benchmarks using BETR achieves a 2.1x compute multiplier over DCLM-Baseline (4.7x over unfiltered data) and improves performance on 9 out of 10 tasks across all scales. BETR also generalizes well: when targeting a diverse set of benchmarks disjoint from our evaluation suite, it still matches or outperforms baselines. Our scaling analysis further reveals a clear trend: larger models require less aggressive filtering. Overall, our findings show that directly matching pretraining data to target tasks precisely shapes model capabilities and highlight that optimal selection strategies must adapt to model scale.
comment: 44 pages, 25 figures, 13 tables
☆ S2WTM: Spherical Sliced-Wasserstein Autoencoder for Topic Modeling ACL 2025
Modeling latent representations in a hyperspherical space has proven effective for capturing directional similarities in high-dimensional text data, benefiting topic modeling. Variational autoencoder-based neural topic models (VAE-NTMs) commonly adopt the von Mises-Fisher prior to encode hyperspherical structure. However, VAE-NTMs often suffer from posterior collapse, where the KL divergence term in the objective function highly diminishes, leading to ineffective latent representations. To mitigate this issue while modeling hyperspherical structure in the latent space, we propose the Spherical Sliced Wasserstein Autoencoder for Topic Modeling (S2WTM). S2WTM employs a prior distribution supported on the unit hypersphere and leverages the Spherical Sliced-Wasserstein distance to align the aggregated posterior distribution with the prior. Experimental results demonstrate that S2WTM outperforms state-of-the-art topic models, generating more coherent and diverse topics while improving performance on downstream tasks.
comment: Accepted as a long paper for ACL 2025 main conference
☆ Can We Predict Alignment Before Models Finish Thinking? Towards Monitoring Misaligned Reasoning Models
Open-weights reasoning language models generate long chains-of-thought (CoTs) before producing a final response, which improves performance but introduces additional alignment risks, with harmful content often appearing in both the CoTs and the final outputs. In this work, we investigate if we can use CoTs to predict final response misalignment. We evaluate a range of monitoring approaches, including humans, highly-capable large language models, and text classifiers, using either CoT text or activations. First, we find that a simple linear probe trained on CoT activations can significantly outperform all text-based methods in predicting whether a final response will be safe or unsafe. CoT texts are often unfaithful and can mislead humans and classifiers, while model latents (i.e., CoT activations) offer a more reliable predictive signal. Second, the probe makes accurate predictions before reasoning completes, achieving strong performance even when applied to early CoT segments. These findings generalize across model sizes, families, and safety benchmarks, suggesting that lightweight probes could enable real-time safety monitoring and early intervention during generation.
☆ Advancing Retrieval-Augmented Generation for Structured Enterprise and Internal Data
Organizations increasingly rely on proprietary enterprise data, including HR records, structured reports, and tabular documents, for critical decision-making. While Large Language Models (LLMs) have strong generative capabilities, they are limited by static pretraining, short context windows, and challenges in processing heterogeneous data formats. Conventional Retrieval-Augmented Generation (RAG) frameworks address some of these gaps but often struggle with structured and semi-structured data. This work proposes an advanced RAG framework that combines hybrid retrieval strategies using dense embeddings (all-mpnet-base-v2) and BM25, enhanced by metadata-aware filtering with SpaCy NER and cross-encoder reranking. The framework applies semantic chunking to maintain textual coherence and retains tabular data structures to preserve row-column integrity. Quantized indexing optimizes retrieval efficiency, while human-in-the-loop feedback and conversation memory improve adaptability. Experiments on enterprise datasets show notable improvements: Precision@5 increased by 15 percent (90 versus 75), Recall@5 by 13 percent (87 versus 74), and Mean Reciprocal Rank by 16 percent (0.85 versus 0.69). Qualitative evaluations show higher scores in Faithfulness (4.6 versus 3.0), Completeness (4.2 versus 2.5), and Relevance (4.5 versus 3.2) on a 5-point Likert scale. These results demonstrate the framework's effectiveness in delivering accurate, comprehensive, and contextually relevant responses for enterprise tasks. Future work includes extending to multimodal data and integrating agent-based retrieval. The source code will be released at https://github.com/CheerlaChandana/Enterprise-Chatbot
☆ Probing for Arithmetic Errors in Language Models
We investigate whether internal activations in language models can be used to detect arithmetic errors. Starting with a controlled setting of 3-digit addition, we show that simple probes can accurately decode both the model's predicted output and the correct answer from hidden states, regardless of whether the model's output is correct. Building on this, we train lightweight error detectors that predict model correctness with over 90% accuracy. We then extend our analysis to structured chain-of-thought traces on addition-only GSM8K problems and find that probes trained on simple arithmetic generalize well to this more complex setting, revealing consistent internal representations. Finally, we demonstrate that these probes can guide selective re-prompting of erroneous reasoning steps, improving task accuracy with minimal disruption to correct outputs. Our findings suggest that arithmetic errors can be anticipated from internal activations alone, and that simple probes offer a viable path toward lightweight model self-correction.
☆ Developing Visual Augmented Q&A System using Scalable Vision Embedding Retrieval & Late Interaction Re-ranker SIGIR
Traditional information extraction systems face challenges with text only language models as it does not consider infographics (visual elements of information) such as tables, charts, images etc. often used to convey complex information to readers. Multimodal LLM (MLLM) face challenges of finding needle in the haystack problem i.e., either longer context length or substantial number of documents as search space. Late interaction mechanism over visual language models has shown state of the art performance in retrieval-based vision augmented Q&A tasks. There are yet few challenges using it for RAG based multi-modal Q&A. Firstly, many popular and widely adopted vector databases do not support native multi-vector retrieval. Secondly, late interaction requires computation which inflates space footprint and can hinder enterprise adoption. Lastly, the current state of late interaction mechanism does not leverage the approximate neighbor search indexing methods for large speed ups in retrieval process. This paper explores a pragmatic approach to make vision retrieval process scalable and efficient without compromising on performance quality. We propose multi-step custom implementation utilizing widely adopted hybrid search (metadata & embedding) and state of the art late interaction re-ranker to retrieve best matching pages. Finally, MLLM are prompted as reader to generate answers from contextualized best matching pages. Through experiments, we observe that the proposed design is scalable (significant speed up) and stable (without degrading performance quality), hence can be used as production systems at enterprises.
comment: Presented at NLP@IR workshop at SIGIR conference
☆ Web-Browsing LLMs Can Access Social Media Profiles and Infer User Demographics
Large language models (LLMs) have traditionally relied on static training data, limiting their knowledge to fixed snapshots. Recent advancements, however, have equipped LLMs with web browsing capabilities, enabling real time information retrieval and multi step reasoning over live web content. While prior studies have demonstrated LLMs ability to access and analyze websites, their capacity to directly retrieve and analyze social media data remains unexplored. Here, we evaluate whether web browsing LLMs can infer demographic attributes of social media users given only their usernames. Using a synthetic dataset of 48 X (Twitter) accounts and a survey dataset of 1,384 international participants, we show that these models can access social media content and predict user demographics with reasonable accuracy. Analysis of the synthetic dataset further reveals how LLMs parse and interpret social media profiles, which may introduce gender and political biases against accounts with minimal activity. While this capability holds promise for computational social science in the post API era, it also raises risks of misuse particularly in information operations and targeted advertising underscoring the need for safeguards. We recommend that LLM providers restrict this capability in public facing applications, while preserving controlled access for verified research purposes.
☆ Beyond Single Models: Enhancing LLM Detection of Ambiguity in Requests through Debate
Large Language Models (LLMs) have demonstrated significant capabilities in understanding and generating human language, contributing to more natural interactions with complex systems. However, they face challenges such as ambiguity in user requests processed by LLMs. To address these challenges, this paper introduces and evaluates a multi-agent debate framework designed to enhance detection and resolution capabilities beyond single models. The framework consists of three LLM architectures (Llama3-8B, Gemma2-9B, and Mistral-7B variants) and a dataset with diverse ambiguities. The debate framework markedly enhanced the performance of Llama3-8B and Mistral-7B variants over their individual baselines, with Mistral-7B-led debates achieving a notable 76.7% success rate and proving particularly effective for complex ambiguities and efficient consensus. While acknowledging varying model responses to collaborative strategies, these findings underscore the debate framework's value as a targeted method for augmenting LLM capabilities. This work offers important insights for developing more robust and adaptive language understanding systems by showing how structured debates can lead to improved clarity in interactive systems.
comment: Accepted at the 2025 SICE Festival with Annual Conference (SICE FES)
☆ Exploring Gender Bias in Alzheimer's Disease Detection: Insights from Mandarin and Greek Speech Perception
Gender bias has been widely observed in speech perception tasks, influenced by the fundamental voicing differences between genders. This study reveals a gender bias in the perception of Alzheimer's Disease (AD) speech. In a perception experiment involving 16 Chinese listeners evaluating both Chinese and Greek speech, we identified that male speech was more frequently identified as AD, with this bias being particularly pronounced in Chinese speech. Acoustic analysis showed that shimmer values in male speech were significantly associated with AD perception, while speech portion exhibited a significant negative correlation with AD identification. Although language did not have a significant impact on AD perception, our findings underscore the critical role of gender bias in AD speech perception. This work highlights the necessity of addressing gender bias when developing AD detection models and calls for further research to validate model performance across different linguistic contexts.
comment: 12 pages, 5 figures, conference or other essential info
☆ Nonlinear Concept Erasure: a Density Matching Approach ECAI 2025
Ensuring that neural models used in real-world applications cannot infer sensitive information, such as demographic attributes like gender or race, from text representations is a critical challenge when fairness is a concern. We address this issue through concept erasure, a process that removes information related to a specific concept from distributed representations while preserving as much of the remaining semantic information as possible. Our approach involves learning an orthogonal projection in the embedding space, designed to make the class-conditional feature distributions of the discrete concept to erase indistinguishable after projection. By adjusting the rank of the projector, we control the extent of information removal, while its orthogonality ensures strict preservation of the local structure of the embeddings. Our method, termed $\overline{\mathrm{L}}$EOPARD, achieves state-of-the-art performance in nonlinear erasure of a discrete attribute on classic natural language processing benchmarks. Furthermore, we demonstrate that $\overline{\mathrm{L}}$EOPARD effectively mitigates bias in deep nonlinear classifiers, thereby promoting fairness.
comment: 17 pages, 10 figures, accepted for publication in ECAI 2025 (28th European Conference on Artificial Intelligence)
☆ Chain-of-Descriptions: Improving Code LLMs for VHDL Code Generation and Summarization
Large Language Models (LLMs) have become widely used across diverse NLP tasks and domains, demonstrating their adaptability and effectiveness. In the realm of Electronic Design Automation (EDA), LLMs show promise for tasks like Register-Transfer Level (RTL) code generation and summarization. However, despite the proliferation of LLMs for general code-related tasks, there's a dearth of research focused on evaluating and refining these models for hardware description languages (HDLs), notably VHDL. In this study, we evaluate the performance of existing code LLMs for VHDL code generation and summarization using various metrics and two datasets -- VHDL-Eval and VHDL-Xform. The latter, an in-house dataset, aims to gauge LLMs' understanding of functionally equivalent code. Our findings reveal consistent underperformance of these models across different metrics, underscoring a significant gap in their suitability for this domain. To address this challenge, we propose Chain-of-Descriptions (CoDes), a novel approach to enhance the performance of LLMs for VHDL code generation and summarization tasks. CoDes involves generating a series of intermediate descriptive steps based on: (i) the problem statement for code generation, and (ii) the VHDL code for summarization. These steps are then integrated with the original input prompt (problem statement or code) and provided as input to the LLMs to generate the final output. Our experiments demonstrate that the CoDes approach significantly surpasses the standard prompting strategy across various metrics on both datasets. This method not only improves the quality of VHDL code generation and summarization but also serves as a framework for future research aimed at enhancing code LLMs for VHDL.
comment: 10 pages (6 content pages + 4 supplementary), 5 figures, Proceedings of the 2024 ACM/IEEE International Symposium on Machine Learning for CAD. 2024 (MLCAD'24)
☆ Text-ADBench: Text Anomaly Detection Benchmark based on LLMs Embedding
Text anomaly detection is a critical task in natural language processing (NLP), with applications spanning fraud detection, misinformation identification, spam detection and content moderation, etc. Despite significant advances in large language models (LLMs) and anomaly detection algorithms, the absence of standardized and comprehensive benchmarks for evaluating the existing anomaly detection methods on text data limits rigorous comparison and development of innovative approaches. This work performs a comprehensive empirical study and introduces a benchmark for text anomaly detection, leveraging embeddings from diverse pre-trained language models across a wide array of text datasets. Our work systematically evaluates the effectiveness of embedding-based text anomaly detection by incorporating (1) early language models (GloVe, BERT); (2) multiple LLMs (LLaMa-2, LLama-3, Mistral, OpenAI (small, ada, large)); (3) multi-domain text datasets (news, social media, scientific publications); (4) comprehensive evaluation metrics (AUROC, AUPRC). Our experiments reveal a critical empirical insight: embedding quality significantly governs anomaly detection efficacy, and deep learning-based approaches demonstrate no performance advantage over conventional shallow algorithms (e.g., KNN, Isolation Forest) when leveraging LLM-derived embeddings.In addition, we observe strongly low-rank characteristics in cross-model performance matrices, which enables an efficient strategy for rapid model evaluation (or embedding evaluation) and selection in practical applications. Furthermore, by open-sourcing our benchmark toolkit that includes all embeddings from different models and code at https://github.com/jicongfan/Text-Anomaly-Detection-Benchmark, this work provides a foundation for future research in robust and scalable text anomaly detection systems.
☆ MERA Code: A Unified Framework for Evaluating Code Generation Across Tasks
Advancements in LLMs have enhanced task automation in software engineering; however, current evaluations primarily focus on natural language tasks, overlooking code quality. Most benchmarks prioritize high-level reasoning over executable code and real-world performance, leaving gaps in understanding true capabilities and risks associated with these models in production. To address this issue, we propose MERA Code, a new addition to the MERA benchmark family, specifically focused on evaluating code for the latest code generation LLMs in Russian. This benchmark includes 11 evaluation tasks that span 8 programming languages. Our proposed evaluation methodology features a taxonomy that outlines the practical coding skills necessary for models to complete these tasks. The benchmark comprises an open-source codebase for users to conduct MERA assessments, a scoring system compatible with various programming environments, and a platform featuring a leaderboard and submission system. We evaluate open LLMs and frontier API models, analyzing their limitations in terms of practical coding tasks in non-English languages. We are publicly releasing MERA to guide future research, anticipate groundbreaking features in model development, and standardize evaluation procedures.
☆ Infherno: End-to-end Agent-based FHIR Resource Synthesis from Free-form Clinical Notes EMNLP 2025
For clinical data integration and healthcare services, the HL7 FHIR standard has established itself as a desirable format for interoperability between complex health data. Previous attempts at automating the translation from free-form clinical notes into structured FHIR resources rely on modular, rule-based systems or LLMs with instruction tuning and constrained decoding. Since they frequently suffer from limited generalizability and structural inconformity, we propose an end-to-end framework powered by LLM agents, code execution, and healthcare terminology database tools to address these issues. Our solution, called Infherno, is designed to adhere to the FHIR document schema and competes well with a human baseline in predicting FHIR resources from unstructured text. The implementation features a front end for custom and synthetic data and both local and proprietary models, supporting clinical data integration processes and interoperability across institutions.
comment: Submitted to EMNLP 2025 System Demonstrations | Code: https://github.com/j-frei/Infherno | Video: https://www.youtube.com/watch?v=kyj5C2ivbMw | Demo: https://infherno.misit-augsburg.de | HuggingFace Spaces: https://huggingface.co/spaces/nfel/infherno
☆ Translationese-index: Using Likelihood Ratios for Graded and Generalizable Measurement of Translationese
In this paper, we propose the first quantitative measure for translationese -- the translationese-index (T-index) for graded and generalizable measurement of translationese, computed from the likelihood ratios of two contrastively fine-tuned language models (LMs). We use a synthesized dataset and a dataset with translations in the wild to evaluate T-index's generalizability in cross-domain settings and its validity against human judgments. Our results show that T-index is both robust and efficient. T-index scored by two 0.5B LMs fine-tuned on only 1-5k pairs of synthetic data can well capture translationese in the wild. We find that the relative differences in T-indices between translations can well predict pairwise translationese annotations obtained from human annotators; and the absolute values of T-indices correlate well with human ratings of degrees of translationese (Pearson's $r = 0.568$). Additionally, the correlation between T-index and existing machine translation (MT) quality estimation (QE) metrics such as BLEU and COMET is low, suggesting that T-index is not covered by these metrics and can serve as a complementary metric in MT QE.
☆ Improving Contextual ASR via Multi-grained Fusion with Large Language Models
While end-to-end Automatic Speech Recognition (ASR) models have shown impressive performance in transcribing general speech, they often struggle to accurately recognize contextually relevant keywords, such as proper nouns or user-specific entities. Previous approaches have explored leveraging keyword dictionaries in the textual modality to improve keyword recognition, either through token-level fusion that guides token-by-token generation or phrase-level fusion that enables direct copying of keyword phrases. However, these methods operate at different granularities and have their own limitations. In this paper, we propose a novel multi-grained fusion approach that jointly leverages the strengths of both token-level and phrase-level fusion with Large Language Models (LLMs). Our approach incorporates a late-fusion strategy that elegantly combines ASR's acoustic information with LLM's rich contextual knowledge, balancing fine-grained token precision with holistic phrase-level understanding. Experiments on Chinese and English datasets demonstrate that our approach achieves state-of-the-art performance on keyword-related metrics while preserving high accuracy on non-keyword text. Ablation studies further confirm that the token-level and phrase-level components both contribute significantly to the performance gains, complementing each other in our joint multi-grained framework. The code and models will be publicly available at https://github.com/.
☆ Towards few-shot isolated word reading assessment
We explore an ASR-free method for isolated word reading assessment in low-resource settings. Our few-shot approach compares input child speech to a small set of adult-provided reference templates. Inputs and templates are encoded using intermediate layers from large self-supervised learned (SSL) models. Using an Afrikaans child speech benchmark, we investigate design options such as discretising SSL features and barycentre averaging of the templates. Idealised experiments show reasonable performance for adults, but a substantial drop for child speech input, even with child templates. Despite the success of employing SSL representations in low-resource speech tasks, our work highlights the limitations of SSL representations for processing child data when used in a few-shot classification system.
comment: Accepted to SLaTE 2025
☆ Toward a Behavioural Translation Style Space: Simulating the Temporal Dynamics of Affect, Behaviour, and Cognition in Human Translation Production
The paper introduces a Behavioural Translation Style Space (BTSS) that describes possible behavioural translation patterns. The suggested BTSS is organized as a hierarchical structure that entails various embedded processing layers. We posit that observable translation behaviour - i.e., eye and finger movements - is fundamental when executing the physical act of translation but it is caused and shaped by higher-order cognitive processes and affective translation states. We analyse records of keystrokes and gaze data as indicators of the hidden mental processing structure and organize the behavioural patterns as a multi-layered embedded BTSS. The BTSS serves as the basis for a computational translation agent to simulate the temporal dynamics of affect, automatized behaviour and cognition during human translation production.
☆ RUMAA: Repeat-Aware Unified Music Audio Analysis for Score-Performance Alignment, Transcription, and Mistake Detection SP
This study introduces RUMAA, a transformer-based framework for music performance analysis that unifies score-to-performance alignment, score-informed transcription, and mistake detection in a near end-to-end manner. Unlike prior methods addressing these tasks separately, RUMAA integrates them using pre-trained score and audio encoders and a novel tri-stream decoder capturing task interdependencies through proxy tasks. It aligns human-readable MusicXML scores with repeat symbols to full-length performance audio, overcoming traditional MIDI-based methods that rely on manually unfolded score-MIDI data with pre-specified repeat structures. RUMAA matches state-of-the-art alignment methods on non-repeated scores and outperforms them on scores with repeats in a public piano music dataset, while also delivering promising transcription and mistake detection results.
comment: Accepted to WASPAA 2025
Overview of the Sensemaking Task at the ELOQUENT 2025 Lab: LLMs as Teachers, Students and Evaluators
ELOQUENT is a set of shared tasks that aims to create easily testable high-level criteria for evaluating generative language models. Sensemaking is one such shared task. In Sensemaking, we try to assess how well generative models ``make sense out of a given text'' in three steps inspired by exams in a classroom setting: (1) Teacher systems should prepare a set of questions, (2) Student systems should answer these questions, and (3) Evaluator systems should score these answers, all adhering rather strictly to a given set of input materials. We report on the 2025 edition of Sensemaking, where we had 7 sources of test materials (fact-checking analyses of statements, textbooks, transcribed recordings of a lecture, and educational videos) spanning English, German, Ukrainian, and Czech languages. This year, 4 teams participated, providing us with 2 Teacher submissions, 2 Student submissions, and 2 Evaluator submissions. We added baselines for Teacher and Student using commercial large language model systems. We devised a fully automatic evaluation procedure, which we compare to a minimalistic manual evaluation. We were able to make some interesting observations. For the first task, the creation of questions, better evaluation strategies will still have to be devised because it is difficult to discern the quality of the various candidate question sets. In the second task, question answering, the LLMs examined overall perform acceptably, but restricting their answers to the given input texts remains problematic. In the third task, evaluation of question answers, our adversarial tests reveal that systems using the LLM-as-a-Judge paradigm erroneously rate both garbled question-answer pairs and answers to mixed-up questions as acceptable.
comment: 30 pages, 7 figures, CLEF 2025 Conference and Labs of the Evaluation Forum
☆ RiemannLoRA: A Unified Riemannian Framework for Ambiguity-Free LoRA Optimization
Low-Rank Adaptation (LoRA) has become a widely adopted standard for parameter-efficient fine-tuning of large language models (LLMs), significantly reducing memory and computational demands. However, challenges remain, including finding optimal initialization strategies or mitigating overparametrization in low-rank matrix factorization. In this work, we propose a novel approach that addresses both of the challenges simultaneously within a unified framework. Our method treats a set of fixed-rank LoRA matrices as a smooth manifold. Considering adapters as elements on this manifold removes overparametrization, while determining the direction of the fastest loss decrease along the manifold provides initialization. Special care is taken to obtain numerically stable and computationally efficient implementation of our method, using best practices from numerical linear algebra and Riemannian optimization. Experimental results on LLM and diffusion model architectures demonstrate that RiemannLoRA consistently improves both convergence speed and final performance over standard LoRA and its state-of-the-art modifications.
☆ Iterative Augmentation with Summarization Refinement (IASR) Evaluation for Unstructured Survey data Modeling and Analysis
Text data augmentation is a widely used strategy for mitigating data sparsity in natural language processing (NLP), particularly in low-resource settings where limited samples hinder effective semantic modeling. While augmentation can improve input diversity and downstream interpretability, existing techniques often lack mechanisms to ensure semantic preservation during large-scale or iterative generation, leading to redundancy and instability. This work introduces a principled evaluation framework for large language model (LLM) based text augmentation, comprising two components: (1) Scalability Analysis, which measures semantic consistency as augmentation volume increases, and (2) Iterative Augmentation with Summarization Refinement (IASR), which evaluates semantic drift across recursive paraphrasing cycles. Empirical evaluations across state-of-the-art LLMs show that GPT-3.5 Turbo achieved the best balance of semantic fidelity, diversity, and generation efficiency. Applied to a real-world topic modeling task using BERTopic with GPT-enhanced few-shot labeling, the proposed approach results in a 400% increase in topic granularity and complete elimination of topic overlaps. These findings validated the utility of the proposed frameworks for structured evaluation of LLM-based augmentation in practical NLP pipelines.
☆ Findings of MEGA: Maths Explanation with LLMs using the Socratic Method for Active Learning
This paper presents an intervention study on the effects of the combined methods of (1) the Socratic method, (2) Chain of Thought (CoT) reasoning, (3) simplified gamification and (4) formative feedback on university students' Maths learning driven by large language models (LLMs). We call our approach Mathematics Explanations through Games by AI LLMs (MEGA). Some students struggle with Maths and as a result avoid Math-related discipline or subjects despite the importance of Maths across many fields, including signal processing. Oftentimes, students' Maths difficulties stem from suboptimal pedagogy. We compared the MEGA method to the traditional step-by-step (CoT) method to ascertain which is better by using a within-group design after randomly assigning questions for the participants, who are university students. Samples (n=60) were randomly drawn from each of the two test sets of the Grade School Math 8K (GSM8K) and Mathematics Aptitude Test of Heuristics (MATH) datasets, based on the error margin of 11%, the confidence level of 90%, and a manageable number of samples for the student evaluators. These samples were used to evaluate two capable LLMs at length (Generative Pretrained Transformer 4o (GPT4o) and Claude 3.5 Sonnet) out of the initial six that were tested for capability. The results showed that students agree in more instances that the MEGA method is experienced as better for learning for both datasets. It is even much better than the CoT (47.5% compared to 26.67%) in the more difficult MATH dataset, indicating that MEGA is better at explaining difficult Maths problems.
comment: This paper was accepted for the special issue AI for Education by the IEEE Signal Processing Magazine journal
☆ BOOKCOREF: Coreference Resolution at Book Scale ACL 2025
Coreference Resolution systems are typically evaluated on benchmarks containing small- to medium-scale documents. When it comes to evaluating long texts, however, existing benchmarks, such as LitBank, remain limited in length and do not adequately assess system capabilities at the book scale, i.e., when co-referring mentions span hundreds of thousands of tokens. To fill this gap, we first put forward a novel automatic pipeline that produces high-quality Coreference Resolution annotations on full narrative texts. Then, we adopt this pipeline to create the first book-scale coreference benchmark, BOOKCOREF, with an average document length of more than 200,000 tokens. We carry out a series of experiments showing the robustness of our automatic procedure and demonstrating the value of our resource, which enables current long-document coreference systems to gain up to +20 CoNLL-F1 points when evaluated on full books. Moreover, we report on the new challenges introduced by this unprecedented book-scale setting, highlighting that current models fail to deliver the same performance they achieve on smaller documents. We release our data and code to encourage research and development of new book-scale Coreference Resolution systems at https://github.com/sapienzanlp/bookcoref.
comment: Accepted to ACL 2025 Main Conference. 19 pages
☆ StylOch at PAN: Gradient-Boosted Trees with Frequency-Based Stylometric Features
This submission to the binary AI detection task is based on a modular stylometric pipeline, where: public spaCy models are used for text preprocessing (including tokenisation, named entity recognition, dependency parsing, part-of-speech tagging, and morphology annotation) and extracting several thousand features (frequencies of n-grams of the above linguistic annotations); light-gradient boosting machines are used as the classifier. We collect a large corpus of more than 500 000 machine-generated texts for the classifier's training. We explore several parameter options to increase the classifier's capacity and take advantage of that training set. Our approach follows the non-neural, computationally inexpensive but explainable approach found effective previously.
☆ Evaluating the Ability of Large Language Models to Reason about Cardinal Directions, Revisited IJCAI
We investigate the abilities of 28 Large language Models (LLMs) to reason about cardinal directions (CDs) using a benchmark generated from a set of templates, extensively testing an LLM's ability to determine the correct CD given a particular scenario. The templates allow for a number of degrees of variation such as means of locomotion of the agent involved, and whether set in the first, second or third person. Even the newer Large Reasoning Models are unable to reliably determine the correct CD for all questions. This paper summarises and extends earlier work presented at COSIT-24.
comment: 8 pages, 5 figures. Accepted at QR 2025 : 38th International Workshop on Qualitative Reasoning at IJCAI
☆ A Comparative Approach to Assessing Linguistic Creativity of Large Language Models and Humans
The following paper introduces a general linguistic creativity test for humans and Large Language Models (LLMs). The test consists of various tasks aimed at assessing their ability to generate new original words and phrases based on word formation processes (derivation and compounding) and on metaphorical language use. We administered the test to 24 humans and to an equal number of LLMs, and we automatically evaluated their answers using OCSAI tool for three criteria: Originality, Elaboration, and Flexibility. The results show that LLMs not only outperformed humans in all the assessed criteria, but did better in six out of the eight test tasks. We then computed the uniqueness of the individual answers, which showed some minor differences between humans and LLMs. Finally, we performed a short manual analysis of the dataset, which revealed that humans are more inclined towards E(extending)-creativity, while LLMs favor F(ixed)-creativity.
comment: Accepted for presentation at KES 2025. To appear in Procedia Computer Science (Elsevier)
☆ Improving Data and Parameter Efficiency of Neural Language Models Using Representation Analysis
This thesis addresses challenges related to data and parameter efficiency in neural language models, with a focus on representation analysis and the introduction of new optimization techniques. The first part examines the properties and dynamics of language representations within neural models, emphasizing their significance in enhancing robustness and generalization. It proposes innovative approaches based on representation smoothness, including regularization strategies that utilize Jacobian and Hessian matrices to stabilize training and mitigate sensitivity to input perturbations. The second part focuses on methods to significantly enhance data and parameter efficiency by integrating active learning strategies with parameter-efficient fine-tuning, guided by insights from representation smoothness analysis. It presents smoothness-informed early-stopping techniques designed to eliminate the need for labeled validation sets and proposes innovative combinations of active learning and parameter-efficient fine-tuning to reduce labeling efforts and computational resources. Extensive experimental evaluations across various NLP tasks demonstrate that these combined approaches substantially outperform traditional methods in terms of performance, stability, and efficiency. The third part explores weak supervision techniques enhanced by in-context learning to effectively utilize unlabeled data, further reducing dependence on extensive labeling. It shows that using in-context learning as a mechanism for weak supervision enables models to better generalize from limited labeled data by leveraging unlabeled examples more effectively during training. Comprehensive empirical evaluations confirm significant gains in model accuracy, adaptability, and robustness, especially in low-resource settings and dynamic data environments.
☆ Simplifications are Absolutists: How Simplified Language Reduces Word Sense Awareness in LLM-Generated Definitions
Large Language Models (LLMs) can provide accurate word definitions and explanations for any context. However, the scope of the definition changes for different target groups, like children or language learners. This is especially relevant for homonyms, words with multiple meanings, where oversimplification might risk information loss by omitting key senses, potentially misleading users who trust LLM outputs. We investigate how simplification impacts homonym definition quality across three target groups: Normal, Simple, and ELI5. Using two novel evaluation datasets spanning multiple languages, we test DeepSeek v3, Llama 4 Maverick, Qwen3-30B A3B, GPT-4o mini, and Llama 3.1 8B via LLM-as-Judge and human annotations. Our results show that simplification drastically degrades definition completeness by neglecting polysemy, increasing the risk of misunderstanding. Fine-tuning Llama 3.1 8B with Direct Preference Optimization substantially improves homonym response quality across all prompt types. These findings highlight the need to balance simplicity and completeness in educational NLP to ensure reliable, context-aware definitions for all learners.
comment: Accepted by RANLP 2025
☆ Value-Based Large Language Model Agent Simulation for Mutual Evaluation of Trust and Interpersonal Closeness
Large language models (LLMs) have emerged as powerful tools for simulating complex social phenomena using human-like agents with specific traits. In human societies, value similarity is important for building trust and close relationships; however, it remains unexplored whether this principle holds true in artificial societies comprising LLM agents. Therefore, this study investigates the influence of value similarity on relationship-building among LLM agents through two experiments. First, in a preliminary experiment, we evaluated the controllability of values in LLMs to identify the most effective model and prompt design for controlling the values. Subsequently, in the main experiment, we generated pairs of LLM agents imbued with specific values and analyzed their mutual evaluations of trust and interpersonal closeness following a dialogue. The experiments were conducted in English and Japanese to investigate language dependence. The results confirmed that pairs of agents with higher value similarity exhibited greater mutual trust and interpersonal closeness. Our findings demonstrate that the LLM agent simulation serves as a valid testbed for social science theories, contributes to elucidating the mechanisms by which values influence relationship building, and provides a foundation for inspiring new theories and insights into the social sciences.
☆ Graph Representations for Reading Comprehension Analysis using Large Language Model and Eye-Tracking Biomarker
Reading comprehension is a fundamental skill in human cognitive development. With the advancement of Large Language Models (LLMs), there is a growing need to compare how humans and LLMs understand language across different contexts and apply this understanding to functional tasks such as inference, emotion interpretation, and information retrieval. Our previous work used LLMs and human biomarkers to study the reading comprehension process. The results showed that the biomarkers corresponding to words with high and low relevance to the inference target, as labeled by the LLMs, exhibited distinct patterns, particularly when validated using eye-tracking data. However, focusing solely on individual words limited the depth of understanding, which made the conclusions somewhat simplistic despite their potential significance. This study used an LLM-based AI agent to group words from a reading passage into nodes and edges, forming a graph-based text representation based on semantic meaning and question-oriented prompts. We then compare the distribution of eye fixations on important nodes and edges. Our findings indicate that LLMs exhibit high consistency in language understanding at the level of graph topological structure. These results build on our previous findings and offer insights into effective human-AI co-learning strategies.
☆ Toxicity-Aware Few-Shot Prompting for Low-Resource Singlish Translation
As online communication increasingly incorporates under-represented languages and colloquial dialects, standard translation systems often fail to preserve local slang, code-mixing, and culturally embedded markers of harmful speech. Translating toxic content between low-resource language pairs poses additional challenges due to scarce parallel data and safety filters that sanitize offensive expressions. In this work, we propose a reproducible, two-stage framework for toxicity-preserving translation, demonstrated on a code-mixed Singlish safety corpus. First, we perform human-verified few-shot prompt engineering: we iteratively curate and rank annotator-selected Singlish-target examples to capture nuanced slang, tone, and toxicity. Second, we optimize model-prompt pairs by benchmarking several large language models using semantic similarity via direct and back-translation. Quantitative human evaluation confirms the effectiveness and efficiency of our pipeline. Beyond improving translation quality, our framework contributes to the safety of multicultural LLMs by supporting culturally sensitive moderation and benchmarking in low-resource contexts. By positioning Singlish as a testbed for inclusive NLP, we underscore the importance of preserving sociolinguistic nuance in real-world applications such as content moderation and regional platform governance.
☆ PoTPTQ: A Two-step Power-of-Two Post-training for LLMs ECAI 2025
Large Language Models (LLMs) have demonstrated remarkable performance across various natural language processing (NLP) tasks. However, their deployment is challenging due to the substantial computational resources required. Power-of-two (PoT) quantization is a general tool to counteract this difficulty. Albeit previous works on PoT quantization can be efficiently dequantized on CPUs using fixed-point addition, it showed less effectiveness on GPUs. The reason is entanglement of the sign bit and sequential bit manipulations needed for dequantization. We propose a novel POT quantization framework for LLM weights that (i) outperforms state-of-the-art accuracy in extremely low-precision number formats, and (ii) enables faster inference through more efficient dequantization. To maintain the accuracy of the quantized model, we introduce a two-step post-training algorithm: (i) initialize the quantization scales with a robust starting point, and (ii) refine these scales using a minimal calibration set. The performance of our PoT post-training algorithm surpasses the current state-of-the-art in integer quantization, particularly at low precisions such as 2- and 3-bit formats. Our PoT quantization accelerates the dequantization step required for the floating point inference and leads to $3.67\times$ speed up on a NVIDIA V100, and $1.63\times$ on a NVIDIA RTX 4090, compared to uniform integer dequantization.
comment: Accepted at ECAI 2025 (European Conference on Artificial Intelligence)
☆ The benefits of query-based KGQA systems for complex and temporal questions in LLM era
Large language models excel in question-answering (QA) yet still struggle with multi-hop reasoning and temporal questions. Query-based knowledge graph QA (KGQA) offers a modular alternative by generating executable queries instead of direct answers. We explore multi-stage query-based framework for WikiData QA, proposing multi-stage approach that enhances performance on challenging multi-hop and temporal benchmarks. Through generalization and rejection studies, we evaluate robustness across multi-hop and temporal QA datasets. Additionally, we introduce a novel entity linking and predicate matching method using CoT reasoning. Our results demonstrate the potential of query-based multi-stage KGQA framework for improving multi-hop and temporal QA with small language models. Code and data: https://github.com/ar2max/NLDB-KGQA-System
comment: 15 pages, 3 figures, 7 tables
☆ IAM: Efficient Inference through Attention Mapping between Different-scale LLMs ACL 2025
LLMs encounter significant challenges in resource consumption nowadays, especially with long contexts. Despite extensive efforts dedicate to enhancing inference efficiency, these methods primarily exploit internal sparsity within the models, without leveraging external information for optimization. We identify the high similarity of attention matrices across different-scale LLMs, which offers a novel perspective for optimization. We first conduct a comprehensive analysis of how to measure similarity, how to select mapping Layers and whether mapping is consistency. Based on these insights, we introduce the IAM framework, which achieves dual benefits of accelerated attention computation and reduced KV cache usage by performing attention mapping between small and large LLMs. Our experimental results demonstrate that IAM can accelerate prefill by 15% and reduce KV cache usage by 22.1% without appreciably sacrificing performance. Experiments on different series of models show the generalizability of IAM. Importantly, it is also orthogonal to many existing KV cache optimization methods, making it a versatile addition to the current toolkit for enhancing LLM efficiency.
comment: ACL 2025
☆ DAC: A Dynamic Attention-aware Approach for Task-Agnostic Prompt Compression ACL 2025
Task-agnostic prompt compression leverages the redundancy in natural language to reduce computational overhead and enhance information density within prompts, especially in long-context scenarios. Existing methods predominantly rely on information entropy as the metric to compress lexical units, aiming to achieve minimal information loss. However, these approaches overlook two critical aspects: (i) the importance of attention-critical tokens at the algorithmic level, and (ii) shifts in information entropy during the compression process. Motivated by these challenges, we propose a dynamic attention-aware approach for task-agnostic prompt compression (DAC). This approach effectively integrates entropy and attention information, dynamically sensing entropy shifts during compression to achieve fine-grained prompt compression. Extensive experiments across various domains, including LongBench, GSM8K, and BBH, show that DAC consistently yields robust and substantial improvements across a diverse range of tasks and LLMs, offering compelling evidence of its efficacy.
comment: ACL 2025
☆ BlockBPE: Parallel BPE Tokenization ICML 2025
Tokenization is a critical preprocessing step in large language model pipelines, yet widely-used implementations remain CPU-bound and suboptimal for batch inference workflows on GPU. We present BlockBPE, a parallel GPU implementation of byte-pair encoding (BPE) that achieves near linear-time complexity under realistic assumptions and is optimized for high-throughput, batch inference. Unlike existing Rust-based tokenizers such as HuggingFace Tokenizers or OpenAI's tiktoken-whose runtimes are dominated by Regex pre-tokenization and exhibit $O(n \log n)$ runtime-BlockBPE eliminates the Regex pre-tokenization which leads to small loss in generation quality, but enables highly parallelized token merges within thread blocks, reducing overall complexity to $O(nd)$ where $d \ll n$. On high-batch inference workloads, BlockBPE achieves up to 2x higher throughput than tiktoken and 2.5x over HuggingFace Tokenizers.
comment: ES-FoMo III: 3rd Workshop on Efficient Systems for Foundation Models (ICML 2025)
☆ POLYCHARTQA: Benchmarking Large Vision-Language Models with Multilingual Chart Question Answering
Charts are a universally adopted medium for interpreting and communicating data. However, existing chart understanding benchmarks are predominantly English-centric, limiting their accessibility and applicability to global audiences. In this paper, we present PolyChartQA, the first large-scale multilingual chart question answering benchmark covering 22,606 charts and 26,151 question-answering pairs across 10 diverse languages. PolyChartQA is built using a decoupled pipeline that separates chart data from rendering code, allowing multilingual charts to be flexibly generated by simply translating the data and reusing the code. We leverage state-of-the-art LLM-based translation and enforce rigorous quality control in the pipeline to ensure the linguistic and semantic consistency of the generated multilingual charts. PolyChartQA facilitates systematic evaluation of multilingual chart understanding. Experiments on both open- and closed-source large vision-language models reveal a significant performance gap between English and other languages, especially low-resource ones with non-Latin scripts. This benchmark lays a foundation for advancing globally inclusive vision-language models.
comment: Work in Progress
☆ A Survey of Deep Learning for Geometry Problem Solving
Geometry problem solving is a key area of mathematical reasoning, which is widely involved in many important fields such as education, mathematical ability assessment of artificial intelligence, and multimodal ability assessment. In recent years, the rapid development of deep learning technology, especially the rise of multimodal large language models, has triggered a widespread research boom. This paper provides a survey of the applications of deep learning in geometry problem solving, including (i) a comprehensive summary of the relevant tasks in geometry problem solving; (ii) a thorough review of related deep learning methods; (iii) a detailed analysis of evaluation metrics and methods; and (iv) a critical discussion of the current challenges and future directions that can be explored. Our goal is to provide a comprehensive and practical reference of deep learning for geometry problem solving to promote further developments in this field. We create a continuously updated list of papers on GitHub: https://github.com/majianz/dl4gps.
comment: Work in progress
☆ Marco-Bench-MIF: On Multilingual Instruction-Following Capability of Large Language Models ACL 2025
Instruction-following capability has become a major ability to be evaluated for Large Language Models (LLMs). However, existing datasets, such as IFEval, are either predominantly monolingual and centered on English or simply machine translated to other languages, limiting their applicability in multilingual contexts. In this paper, we present an carefully-curated extension of IFEval to a localized multilingual version named Marco-Bench-MIF, covering 30 languages with varying levels of localization. Our benchmark addresses linguistic constraints (e.g., modifying capitalization requirements for Chinese) and cultural references (e.g., substituting region-specific company names in prompts) via a hybrid pipeline combining translation with verification. Through comprehensive evaluation of 20+ LLMs on our Marco-Bench-MIF, we found that: (1) 25-35% accuracy gap between high/low-resource languages, (2) model scales largely impact performance by 45-60% yet persists script-specific challenges, and (3) machine-translated data underestimates accuracy by7-22% versus localized data. Our analysis identifies challenges in multilingual instruction following, including keyword consistency preservation and compositional constraint adherence across languages. Our Marco-Bench-MIF is available at https://github.com/AIDC-AI/Marco-Bench-MIF.
comment: ACL 2025 Main Conference paper
☆ LLMs Encode Harmfulness and Refusal Separately
LLMs are trained to refuse harmful instructions, but do they truly understand harmfulness beyond just refusing? Prior work has shown that LLMs' refusal behaviors can be mediated by a one-dimensional subspace, i.e., a refusal direction. In this work, we identify a new dimension to analyze safety mechanisms in LLMs, i.e., harmfulness, which is encoded internally as a separate concept from refusal. There exists a harmfulness direction that is distinct from the refusal direction. As causal evidence, steering along the harmfulness direction can lead LLMs to interpret harmless instructions as harmful, but steering along the refusal direction tends to elicit refusal responses directly without reversing the model's judgment on harmfulness. Furthermore, using our identified harmfulness concept, we find that certain jailbreak methods work by reducing the refusal signals without reversing the model's internal belief of harmfulness. We also find that adversarially finetuning models to accept harmful instructions has minimal impact on the model's internal belief of harmfulness. These insights lead to a practical safety application: The model's latent harmfulness representation can serve as an intrinsic safeguard (Latent Guard) for detecting unsafe inputs and reducing over-refusals that is robust to finetuning attacks. For instance, our Latent Guard achieves performance comparable to or better than Llama Guard 3 8B, a dedicated finetuned safeguard model, across different jailbreak methods. Our findings suggest that LLMs' internal understanding of harmfulness is more robust than their refusal decision to diverse input instructions, offering a new perspective to study AI safety
☆ DualReward: A Dynamic Reinforcement Learning Framework for Cloze Tests Distractor Generation CCL 2025
This paper introduces DualReward, a novel reinforcement learning framework for automatic distractor generation in cloze tests. Unlike conventional approaches that rely primarily on supervised learning or static generative models, our method employs a dual reward structure with adaptive scaling that differentiates between human-created gold standard distractors and model-generated candidates. The framework dynamically adjusts reward signal intensity based on model performance and confidence. We evaluate our approach on both passage-level (CLOTH-F) and sentence-level (MCQ) cloze test datasets, demonstrating consistent improvements over state-of-the-art baselines. Experimental results show that our adaptive reward scaling mechanism provides modest but consistent benefits on homogeneous datasets (CLOTH-F) and more substantial improvements (3.48-3.86% in P@1) on diverse, cross-domain data (MCQ), suggesting its particular effectiveness for handling varied question types and domains. Our work offers a flexible framework that effectively balances learning from reliable human examples while exploring novel, high-quality distractors for automated test generation.
comment: Accepted to CCL 2025
☆ COLA-GEC: A Bidirectional Framework for Enhancing Grammatical Acceptability and Error Correction
Grammatical Error Correction (GEC) and grammatical acceptability judgment (COLA) are core tasks in natural language processing, sharing foundational grammatical knowledge yet typically evolving independently. This paper introduces COLA-GEC, a novel bidirectional framework that enhances both tasks through mutual knowledge transfer. First, we augment grammatical acceptability models using GEC datasets, significantly improving their performance across multiple languages. Second, we integrate grammatical acceptability signals into GEC model training via a dynamic loss function, effectively guiding corrections toward grammatically acceptable outputs. Our approach achieves state-of-the-art results on several multilingual benchmarks. Comprehensive error analysis highlights remaining challenges, particularly in punctuation error correction, providing insights for future improvements in grammatical modeling.
comment: Accepted to CLNLP 2025
☆ Cross-Domain Transfer and Few-Shot Learning for Personal Identifiable Information Recognition
Accurate recognition of personally identifiable information (PII) is central to automated text anonymization. This paper investigates the effectiveness of cross-domain model transfer, multi-domain data fusion, and sample-efficient learning for PII recognition. Using annotated corpora from healthcare (I2B2), legal (TAB), and biography (Wikipedia), we evaluate models across four dimensions: in-domain performance, cross-domain transferability, fusion, and few-shot learning. Results show legal-domain data transfers well to biographical texts, while medical domains resist incoming transfer. Fusion benefits are domain-specific, and high-quality recognition is achievable with only 10% of training data in low-specialization domains.
comment: Accepted to CLNLP 2025
☆ Your LLM Knows the Future: Uncovering Its Multi-Token Prediction Potential
Autoregressive language models are constrained by their inherently sequential nature, generating one token at a time. This paradigm limits inference speed and parallelism, especially during later stages of generation when the direction and semantics of text are relatively certain. In this work, we propose a novel framework that leverages the inherent knowledge of vanilla autoregressive language models about future tokens, combining techniques to realize this potential and enable simultaneous prediction of multiple subsequent tokens. Our approach introduces several key innovations: (1) a masked-input formulation where multiple future tokens are jointly predicted from a common prefix; (2) a gated LoRA formulation that preserves the original LLM's functionality, while equipping it for multi-token prediction; (3) a lightweight, learnable sampler module that generates coherent sequences from the predicted future tokens; (4) a set of auxiliary training losses, including a consistency loss, to enhance the coherence and accuracy of jointly generated tokens; and (5) a speculative generation strategy that expands tokens quadratically in the future while maintaining high fidelity. Our method achieves significant speedups through supervised fine-tuning on pretrained models. For example, it generates code and math nearly 5x faster, and improves general chat and knowledge tasks by almost 2.5x. These gains come without any loss in quality.
☆ ILID: Native Script Language Identification for Indian Languages
The language identification task is a crucial fundamental step in NLP. Often it serves as a pre-processing step for widely used NLP applications such as multilingual machine translation, information retrieval, question and answering, and text summarization. The core challenge of language identification lies in distinguishing languages in noisy, short, and code-mixed environments. This becomes even harder in case of diverse Indian languages that exhibit lexical and phonetic similarities, but have distinct differences. Many Indian languages share the same script making the task even more challenging. In this paper, we release a dataset of 230K sentences consisting of English and all 22 official Indian languages labeled with their language identifiers where data in most languages are newly created. We also develop and release robust baseline models using state-of-the-art approaches in machine learning and deep learning that can aid the research in this field. Our baseline models are comparable to the state-of-the-art models for the language identification task.
comment: 8 pages, 1 figure, 7 tables, Paper accepted in RANLP 2025
☆ Tracing Facts or just Copies? A critical investigation of the Competitions of Mechanisms in Large Language Models
This paper presents a reproducibility study examining how Large Language Models (LLMs) manage competing factual and counterfactual information, focusing on the role of attention heads in this process. We attempt to reproduce and reconcile findings from three recent studies by Ortu et al., Yu, Merullo, and Pavlick and McDougall et al. that investigate the competition between model-learned facts and contradictory context information through Mechanistic Interpretability tools. Our study specifically examines the relationship between attention head strength and factual output ratios, evaluates competing hypotheses about attention heads' suppression mechanisms, and investigates the domain specificity of these attention patterns. Our findings suggest that attention heads promoting factual output do so via general copy suppression rather than selective counterfactual suppression, as strengthening them can also inhibit correct facts. Additionally, we show that attention head behavior is domain-dependent, with larger models exhibiting more specialized and category-sensitive patterns.
comment: 18 Pages, 13 figures
☆ Improving Drug Identification in Overdose Death Surveillance using Large Language Models
The rising rate of drug-related deaths in the United States, largely driven by fentanyl, requires timely and accurate surveillance. However, critical overdose data are often buried in free-text coroner reports, leading to delays and information loss when coded into ICD (International Classification of Disease)-10 classifications. Natural language processing (NLP) models may automate and enhance overdose surveillance, but prior applications have been limited. A dataset of 35,433 death records from multiple U.S. jurisdictions in 2020 was used for model training and internal testing. External validation was conducted using a novel separate dataset of 3,335 records from 2023-2024. Multiple NLP approaches were evaluated for classifying specific drug involvement from unstructured death certificate text. These included traditional single- and multi-label classifiers, as well as fine-tuned encoder-only language models such as Bidirectional Encoder Representations from Transformers (BERT) and BioClinicalBERT, and contemporary decoder-only large language models such as Qwen 3 and Llama 3. Model performance was assessed using macro-averaged F1 scores, and 95% confidence intervals were calculated to quantify uncertainty. Fine-tuned BioClinicalBERT models achieved near-perfect performance, with macro F1 scores >=0.998 on the internal test set. External validation confirmed robustness (macro F1=0.966), outperforming conventional machine learning, general-domain BERT models, and various decoder-only large language models. NLP models, particularly fine-tuned clinical variants like BioClinicalBERT, offer a highly accurate and scalable solution for overdose death classification from free-text reports. These methods can significantly accelerate surveillance workflows, overcoming the limitations of manual ICD-10 coding and supporting near real-time detection of emerging substance use trends.
comment: 30 pages, 1 figure, 4 tables, 2 supplemental figures, 4 supplemental tables, submitted to Journal of Forensic Sciences (JFS)
☆ The first open machine translation system for the Chechen language
We introduce the first open-source model for translation between the vulnerable Chechen language and Russian, and the dataset collected to train and evaluate it. We explore fine-tuning capabilities for including a new language into a large language model system for multilingual translation NLLB-200. The BLEU / ChrF++ scores for our model are 8.34 / 34.69 and 20.89 / 44.55 for translation from Russian to Chechen and reverse direction, respectively. The release of the translation models is accompanied by the distribution of parallel words, phrases and sentences corpora and multilingual sentence encoder adapted to the Chechen language.
comment: 7 pages
☆ A Fuzzy Approach to Project Success: Measuring What Matters
This paper introduces a novel approach to project success evaluation by integrating fuzzy logic into an existing construct. Traditional Likert-scale measures often overlook the context-dependent and multifaceted nature of project success. The proposed hierarchical Type-1 Mamdani fuzzy system prioritizes sustained positive impact for end-users, reducing emphasis on secondary outcomes like stakeholder satisfaction and internal project success. This dynamic approach may provide a more accurate measure of project success and could be adaptable to complex evaluations. Future research will focus on empirical testing and broader applications of fuzzy logic in social science.
comment: 3 pages, 1 figure, presented at FUZZ-IEEE 2025
☆ Mono-InternVL-1.5: Towards Cheaper and Faster Monolithic Multimodal Large Language Models
This paper focuses on monolithic Multimodal Large Language Models (MLLMs), which integrate visual encoding and language decoding into a single model. Existing structures and pre-training strategies for monolithic MLLMs often suffer from unstable optimization and catastrophic forgetting. To address these challenges, our key idea is to embed a new visual parameter space into a pre-trained LLM, enabling stable learning of visual knowledge from noisy data via delta tuning. Based on this principle, we first introduce Mono-InternVL, an advanced monolithic MLLM that incorporates a set of visual experts through a multimodal mixture-of-experts architecture. In addition, we design an innovative Endogenous Visual Pre-training (EViP) for Mono-InternVL to maximize its visual capabilities via progressive learning. Mono-InternVL achieves competitive performance against existing MLLMs but also leads to relatively expensive data cost. Therefore, we further present Mono-InternVL-1.5, a cheaper and stronger monolithic MLLM equipped with an improved EViP (EViP++). EViP++ introduces additional visual attention experts to Mono-InternVL-1.5 and re-organizes the pre-training process in an efficient manner. During inference, it includes a fused CUDA kernel to speed up its MoE operations. With these designs, Mono-InternVL-1.5 significantly reduces training and inference costs, while still maintaining competitive performance with Mono-InternVL. To evaluate our approach, we conduct extensive experiments across 15 benchmarks. Results demonstrate that Mono-InternVL outperforms existing monolithic MLLMs on 12 out of 15 benchmarks, e.g., +114-point improvement over Emu3 on OCRBench. Compared to its modular counterpart, i.e., InternVL-1.5, Mono-InternVL-1.5 achieves similar multimodal performance while reducing first-token latency by up to 69%. Code and models are released at https://github.com/OpenGVLab/Mono-InternVL.
☆ Is This Just Fantasy? Language Model Representations Reflect Human Judgments of Event Plausibility
Language models (LMs) are used for a diverse range of tasks, from question answering to writing fantastical stories. In order to reliably accomplish these tasks, LMs must be able to discern the modal category of a sentence (i.e., whether it describes something that is possible, impossible, completely nonsensical, etc.). However, recent studies have called into question the ability of LMs to categorize sentences according to modality (Michaelov et al., 2025; Kauf et al., 2023). In this work, we identify linear representations that discriminate between modal categories within a variety of LMs, or modal difference vectors. Analysis of modal difference vectors reveals that LMs have access to more reliable modal categorization judgments than previously reported. Furthermore, we find that modal difference vectors emerge in a consistent order as models become more competent (i.e., through training steps, layers, and parameter count). Notably, we find that modal difference vectors identified within LM activations can be used to model fine-grained human categorization behavior. This potentially provides a novel view into how human participants distinguish between modal categories, which we explore by correlating projections along modal difference vectors with human participants' ratings of interpretable features. In summary, we derive new insights into LM modal categorization using techniques from mechanistic interpretability, with the potential to inform our understanding of modal categorization in humans.
☆ Modeling Open-World Cognition as On-Demand Synthesis of Probabilistic Models
When faced with novel situations, people are able to marshal relevant considerations from a wide range of background knowledge and put these to use in inferences and predictions. What permits us to draw in globally relevant information and reason over it coherently? Here, we explore the hypothesis that people use a combination of distributed and symbolic representations to construct bespoke mental models tailored to novel situations. We propose a computational implementation of this idea -- a ``Model Synthesis Architecture'' (MSA) -- using language models to implement global relevance-based retrieval and model synthesis and probabilistic programs to implement bespoke, coherent world models. We evaluate our MSA as a model of human judgments on a novel reasoning dataset. The dataset -- built around a `Model Olympics` domain of sports vignettes -- tests models' capacity for human-like, open-ended reasoning by requiring (i) judgments about novel causal structures described in language; (ii) drawing on large bodies of background knowledge; and (iii) doing both in light of observations that introduce arbitrary novel variables. Our MSA approach captures human judgments better than language model-only baselines, under both direct and chain-of-thought generations from the LM that supports model synthesis. These results suggest that MSAs can be implemented in a way that mirrors people's ability to deliver locally coherent reasoning over globally relevant variables, offering a path to understanding and replicating human reasoning in open-ended domains.
comment: Presented at CogSci 2025
☆ Scaling Up RL: Unlocking Diverse Reasoning in LLMs via Prolonged Training
Recent advancements in reasoning-focused language models such as OpenAI's O1 and DeepSeek-R1 have shown that scaling test-time computation-through chain-of-thought reasoning and iterative exploration-can yield substantial improvements on complex tasks like mathematics and code generation. These breakthroughs have been driven by large-scale reinforcement learning (RL), particularly when combined with verifiable reward signals that provide objective and grounded supervision. In this report, we investigate the effects of prolonged reinforcement learning on a small language model across a diverse set of reasoning domains. Our work identifies several key ingredients for effective training, including the use of verifiable reward tasks, enhancements to Group Relative Policy Optimization (GRPO), and practical techniques to improve training stability and generalization. We introduce controlled KL regularization, clipping ratio, and periodic reference policy resets as critical components for unlocking long-term performance gains. Our model achieves significant improvements over strong baselines, including +14.7% on math, +13.9% on coding, and +54.8% on logic puzzle tasks. To facilitate continued research, we release our model publicly.
comment: 14 pages, 7 figures
☆ Evaluating the Ability of Large Language Models to Reason about Cardinal Directions, Revisited IJCAI
We investigate the abilities of 28 Large language Models (LLMs) to reason about cardinal directions (CDs) using a benchmark generated from a set of templates, extensively testing an LLM's ability to determine the correct CD given a particular scenario. The templates allow for a number of degrees of variation such as means of locomotion of the agent involved, and whether set in the first, second or third person. Even the newer Large Reasoning Models are unable to reliably determine the correct CD for all questions. This paper summarises and extends earlier work presented at COSIT-24.
comment: 8 pages, 5 figures. Accepted at QR 2025 : 38th International Workshop on Qualitative Reasoning at IJCAI. arXiv admin note: substantial text overlap with arXiv:2406.16528
♻ ☆ TD-EVAL: Revisiting Task-Oriented Dialogue Evaluation by Combining Turn-Level Precision with Dialogue-Level Comparisons
Task-oriented dialogue (TOD) systems are experiencing a revolution driven by Large Language Models (LLMs), yet the evaluation methodologies for these systems remain insufficient for their growing sophistication. While traditional automatic metrics effectively assessed earlier modular systems, they focus solely on the dialogue level and cannot detect critical intermediate errors that can arise during user-agent interactions. In this paper, we introduce TD-EVAL (Turn and Dialogue-level Evaluation), a two-step evaluation framework that unifies fine-grained turn-level analysis with holistic dialogue-level comparisons. At turn level, we evaluate each response along three TOD-specific dimensions: conversation cohesion, backend knowledge consistency, and policy compliance. Meanwhile, we design TOD Agent Arena that uses pairwise comparisons to provide a measure of dialogue-level quality. Through experiments on MultiWOZ 2.4 and {\tau}-Bench, we demonstrate that TD-EVAL effectively identifies the conversational errors that conventional metrics miss. Furthermore, TD-EVAL exhibits better alignment with human judgments than traditional and LLM-based metrics. These findings demonstrate that TD-EVAL introduces a new paradigm for TOD system evaluation, efficiently assessing both turn and system levels with a plug-and-play framework for future research.
♻ ☆ Simple Mechanistic Explanations for Out-Of-Context Reasoning ICML 2025
Out-of-context reasoning (OOCR) is a phenomenon in which fine-tuned LLMs exhibit surprisingly deep out-of-distribution generalization. Rather than learning shallow heuristics, they implicitly internalize and act on the consequences of observations scattered throughout the fine-tuning data. In this work, we investigate this phenomenon mechanistically and find that many instances of OOCR in the literature have a simple explanation: the LoRA fine-tuning essentially adds a constant steering vector, steering the model towards a general concept. This improves performance on the fine-tuning task and in many other concept-related domains, causing the surprising generalization. Moreover, we can directly train steering vectors for these tasks from scratch, which also induces OOCR. We find that our results hold even for a task that seems like it must involve conditional behavior (model backdoors); it turns out that unconditionally adding a steering vector is sufficient. Overall, our work presents one explanation of what gets learned during fine-tuning for OOCR tasks, contributing to the key question of why LLMs can reason out of context, an advanced capability that is highly relevant to their safe and reliable deployment.
comment: ICML 2025 Workshop R2-FM
♻ ☆ Towards Agentic RAG with Deep Reasoning: A Survey of RAG-Reasoning Systems in LLMs
Retrieval-Augmented Generation (RAG) lifts the factuality of Large Language Models (LLMs) by injecting external knowledge, yet it falls short on problems that demand multi-step inference; conversely, purely reasoning-oriented approaches often hallucinate or mis-ground facts. This survey synthesizes both strands under a unified reasoning-retrieval perspective. We first map how advanced reasoning optimizes each stage of RAG (Reasoning-Enhanced RAG). Then, we show how retrieved knowledge of different type supply missing premises and expand context for complex inference (RAG-Enhanced Reasoning). Finally, we spotlight emerging Synergized RAG-Reasoning frameworks, where (agentic) LLMs iteratively interleave search and reasoning to achieve state-of-the-art performance across knowledge-intensive benchmarks. We categorize methods, datasets, and open challenges, and outline research avenues toward deeper RAG-Reasoning systems that are more effective, multimodally-adaptive, trustworthy, and human-centric. The collection is available at https://github.com/DavidZWZ/Awesome-RAG-Reasoning.
comment: submitted to ARR May
♻ ☆ Planning-Aware Code Infilling via Horizon-Length Prediction
Fill-in-the-Middle (FIM), or infilling, has become integral to code language models, enabling generation of missing code given both left and right contexts. However, the current FIM training paradigm which performs next-token prediction (NTP) over reordered sequence often leads to models struggling to generate content that aligns well with the surrounding context. We hypothesize that NTP alone is insufficient for models to learn effective planning conditioned on the distant right context, a critical factor for successful code infilling. To overcome this, we propose Horizon-Length Prediction (HLP), a novel training objective that teaches models to predict the number of remaining middle tokens at each step. HLP advances FIM with lookahead planning, enabling models to inherently learn infilling boundaries for arbitrary left and right contexts without relying on dataset-specific post-processing. Our evaluation across different model families and sizes shows that HLP significantly improves FIM performance by up to 24% relatively on diverse benchmarks, across file-level and repository-level. Furthermore, the enhanced planning capability gained through HLP boosts model performance on code reasoning. Importantly, HLP incurs negligible training overhead and no additional inference cost, ensuring its practicality for real-world scenarios.
♻ ☆ From Semantic Web and MAS to Agentic AI: A Unified Narrative of the Web of Agents
The concept of the Web of Agents (WoA), which transforms the static, document-centric Web into an environment of autonomous agents acting on users' behalf, has attracted growing interest as large language models (LLMs) become more capable. However, research in this area is still fragmented across different communities. Contemporary surveys catalog the latest LLM-powered frameworks, while the rich histories of Multi-Agent Systems (MAS) and the Semantic Web are often treated as separate, legacy domains. This fragmentation obscures the intellectual lineage of modern systems and hinders a holistic understanding of the field's trajectory. We present the first comprehensive evolutionary overview of the WoA. We show that modern protocols like A2A and the MCP, are direct evolutionary responses to the well-documented limitations of earlier standards like FIPA standards and OWL-based semantic agents. To systematize this analysis, we introduce a four-axis taxonomy (semantic foundation, communication paradigm, locus of intelligence, discovery mechanism). This framework provides a unified analytical lens for comparing agent architectures across all generations, revealing a clear line of descent where others have seen a disconnect. Our analysis identifies a paradigm shift in the 'locus of intelligence': from being encoded in external data (Semantic Web) or the platform (MAS) to being embedded within the agent's core model (LLM). This shift is foundational to modern Agentic AI, enabling the scalable and adaptive systems the WoA has long envisioned. We conclude that while new protocols are essential, they are insufficient for building a robust, open, trustworthy ecosystem. Finally, we argue that the next research frontier lies in solving persistent socio-technical challenges, and we map out a new agenda focused on decentralized identity, economic models, security, and governance for the emerging WoA.
comment: 33 pages, 9 figures, 8 tables
♻ ☆ Linearly-Interpretable Concept Embedding Models for Text Analysis
Despite their success, Large-Language Models (LLMs) still face criticism due to their lack of interpretability. Traditional post-hoc interpretation methods, based on attention and gradient-based analysis, offer limited insights as they only approximate the model's decision-making processes and have been proved to be unreliable. For this reason, Concept-Bottleneck Models (CBMs) have been lately proposed in the textual field to provide interpretable predictions based on human-understandable concepts. However, CBMs still exhibit several limitations due to their architectural constraints limiting their expressivity, to the absence of task-interpretability when employing non-linear task predictors and for requiring extensive annotations that are impractical for real-world text data. In this paper, we address these challenges by proposing a novel Linearly Interpretable Concept Embedding Model (LICEM) going beyond the current accuracy-interpretability trade-off. LICEMs classification accuracy is better than existing interpretable models and matches black-box ones. We show that the explanations provided by our models are more interveneable and causally consistent with respect to existing solutions. Finally, we show that LICEMs can be trained without requiring any concept supervision, as concepts can be automatically predicted when using an LLM backbone.
♻ ☆ Automated Novelty Evaluation of Academic Paper: A Collaborative Approach Integrating Human and Large Language Model Knowledge
Novelty is a crucial criterion in the peer review process for evaluating academic papers. Traditionally, it's judged by experts or measure by unique reference combinations. Both methods have limitations: experts have limited knowledge, and the effectiveness of the combination method is uncertain. Moreover, it's unclear if unique citations truly measure novelty. The large language model (LLM) possesses a wealth of knowledge, while human experts possess judgment abilities that the LLM does not possess. Therefore, our research integrates the knowledge and abilities of LLM and human experts to address the limitations of novelty assessment. One of the most common types of novelty in academic papers is the introduction of new methods. In this paper, we propose leveraging human knowledge and LLM to assist pretrained language models (PLMs, e.g. BERT etc.) in predicting the method novelty of papers. Specifically, we extract sentences related to the novelty of the academic paper from peer review reports and use LLM to summarize the methodology section of the academic paper, which are then used to fine-tune PLMs. In addition, we have designed a text-guided fusion module with novel Sparse-Attention to better integrate human and LLM knowledge. We compared the method we proposed with a large number of baselines. Extensive experiments demonstrate that our method achieves superior performance.
comment: Journal of the Association for Information Science and Technology, 2025
♻ ☆ NLP Meets the World: Toward Improving Conversations With the Public About Natural Language Processing Research
Recent developments in large language models (LLMs) have been accompanied by rapidly growing public interest in natural language processing (NLP). This attention is reflected by major news venues, which sometimes invite NLP researchers to share their knowledge and views with a wide audience. Recognizing the opportunities of the present, for both the research field and for individual researchers, this paper shares recommendations for communicating with a general audience about the capabilities and limitations of NLP. These recommendations cover three themes: vague terminology as an obstacle to public understanding, unreasonable expectations as obstacles to sustainable growth, and ethical failures as obstacles to continued support. Published NLP research and popular news coverage are cited to illustrate these themes with examples. The recommendations promote effective, transparent communication with the general public about NLP, in order to strengthen public understanding and encourage support for research.
comment: 7 pages
♻ ☆ Measuring Spiritual Values and Bias of Large Language Models
Large language models (LLMs) have become integral tool for users from various backgrounds. LLMs, trained on vast corpora, reflect the linguistic and cultural nuances embedded in their pre-training data. However, the values and perspectives inherent in this data can influence the behavior of LLMs, leading to potential biases. As a result, the use of LLMs in contexts involving spiritual or moral values necessitates careful consideration of these underlying biases. Our work starts with verification of our hypothesis by testing the spiritual values of popular LLMs. Experimental results show that LLMs' spiritual values are quite diverse, as opposed to the stereotype of atheists or secularists. We then investigate how different spiritual values affect LLMs in social-fairness scenarios e.g., hate speech identification). Our findings reveal that different spiritual values indeed lead to different sensitivity to different hate target groups. Furthermore, we propose to continue pre-training LLMs on spiritual texts, and empirical results demonstrate the effectiveness of this approach in mitigating spiritual bias.
comment: 9 pages including appendix; 5 figures; 5 tables
♻ ☆ Hallucination Detox: Sensitivity Dropout (SenD) for Large Language Model Training ACL 2025
As large language models (LLMs) become increasingly prevalent, concerns about their reliability, particularly due to hallucinations - factually inaccurate or irrelevant outputs - have grown. Our research investigates the relationship between the uncertainty in training dynamics and the emergence of hallucinations. Using models from the Pythia suite and several hallucination detection metrics, we analyze hallucination trends and identify significant variance during training. To address this, we propose \textbf{Sensitivity Dropout (SenD)}, a novel training protocol designed to reduce hallucination variance during training by deterministically dropping embedding indices with significant variability. In addition, we develop an unsupervised hallucination detection metric, Efficient EigenScore (EES), which approximates the traditional EigenScore in 2x speed. This metric is integrated into our training protocol, allowing SenD to be both computationally scalable and effective at reducing hallucination variance. SenD improves test-time reliability of Pythia and Meta's Llama models by up to 17\% and enhances factual accuracy in Wikipedia, Medical, Legal, and Coding domains without affecting downstream task performance.
comment: Accepted to ACL 2025, accepted to Safe Generative AI Workshop @ NeurIPS 2024. Camera-ready version for ACL 2025 (to appear). Submitted July 2025
♻ ☆ FADE: Why Bad Descriptions Happen to Good Features
Recent advances in mechanistic interpretability have highlighted the potential of automating interpretability pipelines in analyzing the latent representations within LLMs. While this may enhance our understanding of internal mechanisms, the field lacks standardized evaluation methods for assessing the validity of discovered features. We attempt to bridge this gap by introducing FADE: Feature Alignment to Description Evaluation, a scalable model-agnostic framework for automatically evaluating feature-to-description alignment. FADE evaluates alignment across four key metrics - Clarity, Responsiveness, Purity, and Faithfulness - and systematically quantifies the causes of the misalignment between features and their descriptions. We apply FADE to analyze existing open-source feature descriptions and assess key components of automated interpretability pipelines, aiming to enhance the quality of descriptions. Our findings highlight fundamental challenges in generating feature descriptions, particularly for SAEs compared to MLP neurons, providing insights into the limitations and future directions of automated interpretability. We release FADE as an open-source package at: https://github.com/brunibrun/FADE
♻ ☆ Semantic Adapter for Universal Text Embeddings: Diagnosing and Mitigating Negation Blindness to Enhance Universality ECAI 2025
Negation plays an important role in various natural language processing tasks such as Natural Language Inference and Sentiment Analysis tasks. Numerous prior studies have found that contextual text embedding models such as BERT, ELMO, RoBERTa or XLNet face challenges in accurately understanding negation. Recent advancements in universal text embeddings have demonstrated superior performance over contextual text embeddings in various tasks. However, due to the bias in popular evaluation benchmarks, the negation awareness capacity of these models remains unclear. To bridge the gap in existing literature, an in-depth analysis is initiated in this work to study the negation awareness of cutting-edge universal text embedding models. Our findings reveal a significant lack of negation awareness in these models, often interpreting negated text pairs as semantically similar. To efficiently deal with the conflict that different tasks need different trade-offs between topic and negation information among other semantic information, a data-efficient and computational-efficient embedding re-weighting method is proposed without modifying the parameters of text embedding models. The proposed solution is able to improve text embedding models' negation awareness significantly on both simple negation understanding task and complex negation understanding task. Furthermore, the proposed solution can also significantly improve the negation awareness of Large Language Model based task-specific high dimensional universal text embeddings.
comment: Accepted in ECAI 2025 main track
♻ ☆ Truth Sleuth and Trend Bender: AI Agents to fact-check YouTube videos and influence opinions
Misinformation poses a significant threat in today's digital world, often spreading rapidly through platforms like YouTube. This paper introduces a novel approach to combating misinformation by developing an AI-powered system that not only fact-checks claims made in YouTube videos but also actively engages users in the comment section and challenge misleading narratives. Our system comprises two main agents: Truth Sleuth and Trend Bender. Truth Sleuth extracts claims from a YouTube video, uses a Retrieval-Augmented Generation (RAG) approach - drawing on sources like Wikipedia, Google Search, Google FactCheck - to accurately assess their veracity and generates a nuanced and comprehensive report. Through rigorous prompt engineering, Trend Bender leverages this report along with a curated corpus of relevant articles to generate insightful and persuasive comments designed to stimulate a productive debate. With a carefully set up self-evaluation loop, this agent is able to iteratively improve its style and refine its output. We demonstrate the system's capabilities through experiments on established benchmark datasets and a real-world deployment on YouTube, showcasing its potential to engage users and potentially influence perspectives. Our findings highlight the high accuracy of our fact-checking agent, and confirm the potential of AI-driven interventions in combating misinformation and fostering a more informed online space.
♻ ☆ Reasoning Strategies in Large Language Models: Can They Follow, Prefer, and Optimize?
Human reasoning involves different strategies, each suited to specific problems. Prior work shows that large language model (LLMs) tend to favor a single reasoning strategy, potentially limiting their effectiveness in diverse reasoning challenges. In this work, we investigate whether prompting can control LLMs reasoning strategies and assess its impact on logical problem-solving. While our experiments show that no single strategy consistently improves accuracy, performance could be enhanced if models could adaptively choose the optimal strategy. We propose methods to guide LLMs in strategy selection, highlighting new ways to refine their reasoning abilities.
♻ ☆ TRIM: Token Reduction and Inference Modeling for Cost-Effective Language Generation
The inference cost of Large Language Models (LLMs) is a significant challenge due to their computational demands, specially on tasks requiring long outputs. However, natural language often contains redundancy, which presents an opportunity for optimization. We have observed that LLMs can generate distilled language-concise outputs that retain essential meaning, when prompted appropriately. We propose TRIM, a pipeline for saving computational cost in which a shorter distilled output from the LLM is reconstructed into a full narrative by a smaller model with lower inference costs. Our experiments show promising results, particularly in general knowledge domains with 20.58% saved tokens on average with tiny decrease in evaluation metrics, hinting that this approach can effectively balance efficiency and accuracy in language processing tasks.
comment: 13 pages, 12 tables, 7 figures
♻ ☆ Protecting Copyrighted Material with Unique Identifiers in Large Language Model Training
A primary concern regarding training large language models (LLMs) is whether they abuse copyrighted online text. With the increasing training data scale and the prevalence of LLMs in daily lives, two problems arise: \textbf{1)} false positive membership inference results misled by similar examples; \textbf{2)} membership inference methods are usually too complex for end users to understand and use. To address these issues, we propose an alternative \textit{insert-and-detect} methodology, advocating that web users and content platforms employ \textbf{\textit{unique identifiers}} for reliable and independent membership inference. Users and platforms can create their identifiers, embed them in copyrighted text, and independently detect them in future LLMs. As an initial demonstration, we introduce \textit{\textbf{ghost sentences}} and a user-friendly last-$k$ words test, allowing end users to chat with LLMs for membership inference. Ghost sentences consist primarily of unique passphrases of random natural words, which can come with customized elements to bypass possible filter rules. The last-$k$ words test requires a significant repetition time of ghost sentences~($\ge10$). For cases with fewer repetitions, we designed an extra perplexity test, as LLMs exhibit high perplexity when encountering unnatural passphrases. We also conduct a comprehensive study on the memorization and membership inference of ghost sentences, examining factors such as training data scales, model sizes, repetition times, insertion positions, wordlist of passphrases, alignment, \textit{etc}. Our study shows the possibility of applying ghost sentences in real scenarios and provides instructions for the potential application.
comment: A technical report, work mainly done in the early of 2024
♻ ☆ A Survey of Frontiers in LLM Reasoning: Inference Scaling, Learning to Reason, and Agentic Systems
Reasoning is a fundamental cognitive process that enables logical inference, problem-solving, and decision-making. With the rapid advancement of large language models (LLMs), reasoning has emerged as a key capability that distinguishes advanced AI systems from conventional models that empower chatbots. In this survey, we categorize existing methods along two orthogonal dimensions: (1) Regimes, which define the stage at which reasoning is achieved (either at inference time or through dedicated training); and (2) Architectures, which determine the components involved in the reasoning process, distinguishing between standalone LLMs and agentic compound systems that incorporate external tools, and multi-agent collaborations. Within each dimension, we analyze two key perspectives: (1) Input level, which focuses on techniques that construct high-quality prompts that the LLM condition on; and (2) Output level, which methods that refine multiple sampled candidates to enhance reasoning quality. This categorization provides a systematic understanding of the evolving landscape of LLM reasoning, highlighting emerging trends such as the shift from inference-scaling to learning-to-reason (e.g., DeepSeek-R1), and the transition to agentic workflows (e.g., OpenAI Deep Research, Manus Agent). Additionally, we cover a broad spectrum of learning algorithms, from supervised fine-tuning to reinforcement learning such as PPO and GRPO, and the training of reasoners and verifiers. We also examine key designs of agentic workflows, from established patterns like generator-evaluator and LLM debate to recent innovations. ...
comment: 72 pages, 6 figures
♻ ☆ Large Language Models Often Know When They Are Being Evaluated
If AI models can detect when they are being evaluated, the effectiveness of evaluations might be compromised. For example, models could have systematically different behavior during evaluations, leading to less reliable benchmarks for deployment and governance decisions. We investigate whether frontier language models can accurately classify transcripts based on whether they originate from evaluations or real-world deployment, a capability we call evaluation awareness. To achieve this, we construct a diverse benchmark of 1,000 prompts and transcripts from 61 distinct datasets. These span public benchmarks (e.g., MMLU, SWEBench), real-world deployment interactions, and agent trajectories from scaffolding frameworks (e.g., web-browsing agents). Frontier models clearly demonstrate above-random evaluation awareness (Gemini-2.5-Pro reaches an AUC of $0.83$), but do not yet surpass our simple human baseline (AUC of $0.92$). Furthermore, both AI models and humans are better at identifying evaluations in agentic settings compared to chat settings. Additionally, we test whether models can identify the purpose of the evaluation. Under multiple-choice and open-ended questioning, AI models far outperform random chance in identifying what an evaluation is testing for. Our results indicate that frontier models already exhibit a substantial, though not yet superhuman, level of evaluation-awareness. We recommend tracking this capability in future models.
♻ ☆ Learning to Reason at the Frontier of Learnability
Reinforcement learning is now widely adopted as the final stage of large language model training, especially for reasoning-style tasks such as maths problems. Typically, models attempt each question many times during a single training step and attempt to learn from their successes and failures. However, we demonstrate that throughout training with two popular algorithms (PPO and VinePPO) on two widely used datasets, many questions are either solved by all attempts - meaning they are already learned - or by none - providing no meaningful training signal. To address this, we adapt a method from the reinforcement learning literature - sampling for learnability - and apply it to the reinforcement learning stage of LLM training. Our curriculum prioritises questions with high variance of success, i.e. those where the agent sometimes succeeds, but not always. Our findings demonstrate that this curriculum consistently boosts training performance across multiple algorithms and datasets, paving the way for more efficient and effective reinforcement learning with LLMs.
♻ ☆ RAGGED: Towards Informed Design of Scalable and Stable RAG Systems
Retrieval-augmented generation (RAG) enhances language models by integrating external knowledge, but its effectiveness is highly dependent on system configuration. Improper retrieval settings can degrade performance, making RAG less reliable than closed-book generation. In this work, we introduce RAGGED, a framework for systematically evaluating RAG systems across diverse retriever-reader configurations, retrieval depths, and datasets. Our analysis reveals that reader robustness to noise is the key determinant of RAG stability and scalability. Some readers benefit from increased retrieval depth, while others degrade due to their sensitivity to distracting content. Through large-scale experiments on open-domain, multi-hop, and specialized-domain datasets, we show that retrievers, rerankers, and prompts influence performance but do not fundamentally alter these reader-driven trends. By providing a principled framework and new metrics to assess RAG stability and scalability, RAGGED enables systematic evaluation of retrieval-augmented generation systems, guiding future research on optimizing retrieval depth and model robustness.
comment: Project page: https://github.com/neulab/ragged
♻ ☆ ReviewAgents: Bridging the Gap Between Human and AI-Generated Paper Reviews
Academic paper review is a critical yet time-consuming task within the research community. With the increasing volume of academic publications, automating the review process has become a significant challenge. The primary issue lies in generating comprehensive, accurate, and reasoning-consistent review comments that align with human reviewers' judgments. In this paper, we address this challenge by proposing ReviewAgents, a framework that leverages large language models (LLMs) to generate academic paper reviews. We first introduce a novel dataset, Review-CoT, consisting of 142k review comments, designed for training LLM agents. This dataset emulates the structured reasoning process of human reviewers-summarizing the paper, referencing relevant works, identifying strengths and weaknesses, and generating a review conclusion. Building upon this, we train LLM reviewer agents capable of structured reasoning using a relevant-paper-aware training method. Furthermore, we construct ReviewAgents, a multi-role, multi-LLM agent review framework, to enhance the review comment generation process. Additionally, we propose ReviewBench, a benchmark for evaluating the review comments generated by LLMs. Our experimental results on ReviewBench demonstrate that while existing LLMs exhibit a certain degree of potential for automating the review process, there remains a gap when compared to human-generated reviews. Moreover, our ReviewAgents framework further narrows this gap, outperforming advanced LLMs in generating review comments.
comment: Work in progress
♻ ☆ Labels Generated by Large Language Models Help Measure People's Empathy in Vitro
Large language models (LLMs) have revolutionised many fields, with LLM-as-a-service (LLMSaaS) offering accessible, general-purpose solutions without costly task-specific training. In contrast to the widely studied prompt engineering for directly solving tasks (in vivo), this paper explores LLMs' potential for in-vitro applications: using LLM-generated labels to improve supervised training of mainstream models. We examine two strategies - (1) noisy label correction and (2) training data augmentation - in empathy computing, an emerging task to predict psychology-based questionnaire outcomes from inputs like textual narratives. Crowdsourced datasets in this domain often suffer from noisy labels that misrepresent underlying empathy. We show that replacing or supplementing these crowdsourced labels with LLM-generated labels, developed using psychology-based scale-aware prompts, achieves statistically significant accuracy improvements. Notably, the RoBERTa pre-trained language model (PLM) trained with noise-reduced labels yields a state-of-the-art Pearson correlation coefficient of 0.648 on the public NewsEmp benchmarks. This paper further analyses evaluation metric selection and demographic biases to help guide the future development of more equitable empathy computing models. Code and LLM-generated labels are available at https://github.com/hasan-rakibul/LLMPathy.
comment: This work has been submitted to the IEEE for possible publication
♻ ☆ DEEPER Insight into Your User: Directed Persona Refinement for Dynamic Persona Modeling
To advance personalized applications such as recommendation systems and user behavior prediction, recent research increasingly adopts large language models (LLMs) for human -readable persona modeling. In dynamic real -world scenarios, effective persona modeling necessitates leveraging streaming behavior data to continually optimize user personas. However, existing methods -whether regenerating personas or incrementally extending them with new behaviors -often fail to achieve sustained improvements in persona quality or future behavior prediction accuracy. To address this, we propose DEEPER, a novel approach for dynamic persona modeling that enables continual persona optimization. Specifically, we enhance the model's direction -search capability through an iterative reinforcement learning framework, allowing it to automatically identify effective update directions and optimize personas using discrepancies between user behaviors and model predictions. Extensive experiments on dynamic persona modeling involving 4800 users across 10 domains highlight the superior persona optimization capabilities of DEEPER, delivering an impressive 32.2% average reduction in user behavior prediction error over four update rounds -outperforming the best baseline by a remarkable 22.92%.
♻ ☆ Organize the Web: Constructing Domains Enhances Pre-Training Data Curation ICML 2025
Modern language models are trained on large, unstructured datasets consisting of trillions of tokens and obtained by crawling the web. The unstructured nature makes it difficult to reason about their contents and develop systematic approaches to data curation. In this paper, we unpack monolithic web corpora by developing taxonomies of their contents and organizing them into domains. We introduce WebOrganizer, a framework for organizing web pages in terms of both their topic and format. Using these two complementary notions of domains, we automatically annotate pre-training data by distilling annotations from a large language model into efficient classifiers. This allows us to study how data from different domains should be mixed to improve models on downstream tasks, and we show that we can combine insights about effective topics and formats to further boost performance. We demonstrate that our domain mixing also improves existing methods that select data based on quality. Furthermore, we study and compare how quality-based methods will implicitly change the domain mixture. Overall, our work demonstrates that constructing and mixing domains provides a valuable complement to quality-based data curation methods, opening new avenues for effective and insightful pre-training data curation.
comment: Accepted at ICML 2025. Project page: https://weborganizer.allen.ai
♻ ☆ CultureCLIP: Empowering CLIP with Cultural Awareness through Synthetic Images and Contextualized Captions
Pretrained vision-language models (VLMs) such as CLIP excel in general multimodal comprehension but often struggle to capture nuanced, context-dependent visual cues. This makes it difficult to distinguish between similar-looking concepts with potentially different cultural meanings. Such deficiencies are mainly due to a limited amount of high-quality cultural data, contextual information, and the lack of negative examples that highlight subtle differences. To mitigate this, we design a data curation pipeline leveraging open-sourced VLMs and text-to-image models to construct CulTwin, a synthetic cultural dataset. This dataset consists of paired concept-caption-image triplets, where concepts visually resemble each other but are culturally different. Then, we fine-tune CLIP on CulTwin to develop CultureCLIP, which aligns cultural concepts with contextually enhanced captions and synthetic images through tailored contrastive learning. Experiments on culture-specific benchmarks show that CultureCLIP outperforms the base CLIP, achieving up to a notable 5.49% improvement in fine-grained concept recognition on certain tasks while preserving CLIP's original generalization ability, validating the effectiveness of our data synthesis and VLM backbone training paradigm in capturing subtle cultural distinctions.
comment: 25 pages, COLM 2025
♻ ☆ Decoder-Hybrid-Decoder Architecture for Efficient Reasoning with Long Generation
Recent advances in language modeling have demonstrated the effectiveness of State Space Models (SSMs) for efficient sequence modeling. While hybrid architectures such as Samba and the decoder-decoder architecture, YOCO, have shown promising performance gains over Transformers, prior works have not investigated the efficiency potential of representation sharing between SSM layers. In this paper, we introduce the Gated Memory Unit (GMU), a simple yet effective mechanism for efficient memory sharing across layers. We apply it to create SambaY, a decoder-hybrid-decoder architecture that incorporates GMUs in the cross-decoder to share memory readout states from a Samba-based self-decoder. SambaY significantly enhances decoding efficiency, preserves linear pre-filling time complexity, and boosts long-context performance, all while eliminating the need for explicit positional encoding. Through extensive scaling experiments, we demonstrate that our model exhibits a significantly lower irreducible loss compared to a strong YOCO baseline, indicating superior performance scalability under large-scale compute regimes. Our largest model enhanced with Differential Attention, Phi4-mini-Flash-Reasoning, achieves significantly better performance than Phi4-mini-Reasoning on reasoning tasks such as Math500, AIME24/25, and GPQA Diamond without any reinforcement learning, while delivering up to 10x higher decoding throughput on 2K-length prompts with 32K generation length under the vLLM inference framework. We release our training codebase on open-source data at https://github.com/microsoft/ArchScale.
♻ ☆ BRIDGE: Bootstrapping Text to Control Time-Series Generation via Multi-Agent Iterative Optimization and Diffusion Modeling ICML 2025
Time-series Generation (TSG) is a prominent research area with broad applications in simulations, data augmentation, and counterfactual analysis. While existing methods have shown promise in unconditional single-domain TSG, real-world applications demand for cross-domain approaches capable of controlled generation tailored to domain-specific constraints and instance-level requirements. In this paper, we argue that text can provide semantic insights, domain information and instance-specific temporal patterns, to guide and improve TSG. We introduce ``Text-Controlled TSG'', a task focused on generating realistic time series by incorporating textual descriptions. To address data scarcity in this setting, we propose a novel LLM-based Multi-Agent framework that synthesizes diverse, realistic text-to-TS datasets. Furthermore, we introduce BRIDGE, a hybrid text-controlled TSG framework that integrates semantic prototypes with text description for supporting domain-level guidance. This approach achieves state-of-the-art generation fidelity on 11 of 12 datasets, and improves controllability by up to 12% on MSE and 6% MAE compared to no text input generation, highlighting its potential for generating tailored time-series data.
comment: ICML 2025 Main Conference
♻ ☆ Resona: Improving Context Copying in Linear Recurrence Models with Retrieval
Recent shifts in the space of large language model (LLM) research have shown an increasing focus on novel architectures to compete with prototypical Transformer-based models that have long dominated this space. Linear recurrent models have proven to be a viable competitor due to their computational efficiency. However, such models still demonstrate a sizable gap compared to Transformers in terms of in-context learning among other tasks that require recalling information from a context. In this work, we introduce Resona, a simple and scalable framework for augmenting linear recurrent models with retrieval. Resona augments models with the ability to integrate retrieved information from the provided input context, enabling tailored behavior to diverse task requirements. Experiments on a variety of linear recurrent models demonstrate that Resona-augmented models observe significant performance gains on a variety of synthetic as well as real-world natural language tasks, highlighting its ability to act as a general purpose method to improve the in-context learning and language modeling abilities of linear recurrent LLMs.
comment: Comments: Accepted at COLM 2025 (Conference on Learning with Machines)
♻ ☆ Generative Emergent Communication: Large Language Model is a Collective World Model
Large Language Models (LLMs) have demonstrated a remarkable ability to capture extensive world knowledge, yet how this is achieved without direct sensorimotor experience remains a fundamental puzzle. This study proposes a novel theoretical solution by introducing the Collective World Model hypothesis. We argue that an LLM does not learn a world model from scratch; instead, it learns a statistical approximation of a collective world model that is already implicitly encoded in human language through a society-wide process of embodied, interactive sense-making. To formalize this process, we introduce generative emergent communication (Generative EmCom), a framework built on the Collective Predictive Coding (CPC). This framework models the emergence of language as a process of decentralized Bayesian inference over the internal states of multiple agents. We argue that this process effectively creates an encoder-decoder structure at a societal scale: human society collectively encodes its grounded, internal representations into language, and an LLM subsequently decodes these symbols to reconstruct a latent space that mirrors the structure of the original collective representations. This perspective provides a principled, mathematical explanation for how LLMs acquire their capabilities. The main contributions of this paper are: 1) the formalization of the Generative EmCom framework, clarifying its connection to world models and multi-agent reinforcement learning, and 2) its application to interpret LLMs, explaining phenomena such as distributional semantics as a natural consequence of representation reconstruction. This work provides a unified theory that bridges individual cognitive development, collective language evolution, and the foundations of large-scale AI.
♻ ☆ Learning an Effective Premise Retrieval Model for Efficient Mathematical Formalization
Formalized mathematics has recently garnered significant attention for its ability to assist mathematicians across various fields. Premise retrieval, as a common step in mathematical formalization, has been a challenge, particularly for inexperienced users. Existing retrieval methods that facilitate natural language queries require a certain level of mathematical expertise from users, while approaches based on formal languages (e.g., Lean) typically struggle with the scarcity of training data, hindering the training of effective and generalizable retrieval models. In this work, we introduce a novel method that leverages data extracted from Mathlib to train a lightweight and effective premise retrieval model. In particular, the proposed model embeds queries (i.e., proof state provided by Lean) and premises in a latent space, featuring a tokenizer specifically trained on formal corpora. The model is learned in a contrastive learning framework, in which a fine-grained similarity calculation method and a re-ranking module are applied to enhance the retrieval performance. Experimental results demonstrate that our model outperforms existing baselines, achieving higher accuracy while maintaining a lower computational load. We have released an open-source search engine based on our retrieval model at https://premise-search.com/. The source code and the trained model can be found at https://github.com/ruc-ai4math/Premise-Retrieval.
♻ ☆ Journalism-Guided Agentic In-Context Learning for News Stance Detection
As online news consumption grows, personalized recommendation systems have become integral to digital journalism. However, these systems risk reinforcing filter bubbles and political polarization by failing to incorporate diverse perspectives. Stance detection -- identifying a text's position on a target -- can help mitigate this by enabling viewpoint-aware recommendations and data-driven analyses of media bias. Yet, existing stance detection research remains largely limited to short texts and high-resource languages. To address these gaps, we introduce \textsc{K-News-Stance}, the first Korean dataset for article-level stance detection, comprising 2,000 news articles with article-level and 19,650 segment-level stance annotations across 47 societal issues. We also propose \textsc{JoA-ICL}, a \textbf{Jo}urnalism-guided \textbf{A}gentic \textbf{I}n-\textbf{C}ontext \textbf{L}earning framework that employs a language model agent to predict the stances of key structural segments (e.g., leads, quotes), which are then aggregated to infer the overall article stance. Experiments show that \textsc{JoA-ICL} outperforms existing stance detection methods, highlighting the benefits of segment-level agency in capturing the overall position of long-form news articles. Two case studies further demonstrate its broader utility in promoting viewpoint diversity in news recommendations and uncovering patterns of media bias.
comment: Preprint. 24 pages
♻ ☆ METIS: Fast Quality-Aware RAG Systems with Configuration Adaptation
RAG (Retrieval Augmented Generation) allows LLMs (large language models) to generate better responses with external knowledge, but using more external knowledge often improves generation quality at the expense of response delay. Prior work either reduces the response delay (through better scheduling of RAG queries) or strives to maximize quality (which involves tuning the RAG workflow), but they fall short in optimizing the tradeoff between the delay and quality of RAG responses. This paper presents METIS, the first RAG system that jointly schedules queries and adapts the key RAG configurations of each query, such as the number of retrieved text chunks and synthesis methods, in order to balance quality optimization and response delay reduction. Using 4 popular RAG-QA datasets, we show that compared with the state-of-the-art RAG optimization schemes, METIS reduces the generation latency by $1.64-2.54\times$ without sacrificing generation quality.
comment: 17 pages, 18 figures
♻ ☆ Towards Geo-Culturally Grounded LLM Generations ACL 2025
Generative large language models (LLMs) have demonstrated gaps in diverse cultural awareness across the globe. We investigate the effect of retrieval augmented generation and search-grounding techniques on LLMs' ability to display familiarity with various national cultures. Specifically, we compare the performance of standard LLMs, LLMs augmented with retrievals from a bespoke knowledge base (i.e., KB grounding), and LLMs augmented with retrievals from a web search (i.e., search grounding) on multiple cultural awareness benchmarks. We find that search grounding significantly improves the LLM performance on multiple-choice benchmarks that test propositional knowledge (e.g., cultural norms, artifacts, and institutions), while KB grounding's effectiveness is limited by inadequate knowledge base coverage and a suboptimal retriever. However, search grounding also increases the risk of stereotypical judgments by language models and fails to improve evaluators' judgments of cultural familiarity in a human evaluation with adequate statistical power. These results highlight the distinction between propositional cultural knowledge and open-ended cultural fluency when it comes to evaluating LLMs' cultural awareness.
comment: ACL 2025 (main conference)
♻ ☆ Miipher-2: A Universal Speech Restoration Model for Million-Hour Scale Data Restoration SP
Training data cleaning is a new application for generative model-based speech restoration (SR). This paper introduces Miipher-2, an SR model designed for million-hour scale data, for training data cleaning for large-scale generative models like large language models. Key challenges addressed include generalization to unseen languages, operation without explicit conditioning (e.g., text, speaker ID), and computational efficiency. Miipher-2 utilizes a frozen, pre-trained Universal Speech Model (USM), supporting over 300 languages, as a robust, conditioning-free feature extractor. To optimize efficiency and minimize memory, Miipher-2 incorporates parallel adapters for predicting clean USM features from noisy inputs and employs the WaveFit neural vocoder for waveform synthesis. These components were trained on 3,000 hours of multi-lingual, studio-quality recordings with augmented degradations, while USM parameters remained fixed. Experimental results demonstrate Miipher-2's superior or comparable performance to conventional SR models in word-error-rate, speaker similarity, and both objective and subjective sound quality scores across all tested languages. Miipher-2 operates efficiently on consumer-grade accelerators, achieving a real-time factor of 0.0078, enabling the processing of a million-hour speech dataset in approximately three days using only 100 such accelerators.
comment: Accepted to IEEE WASPAA2025
♻ ☆ UPCORE: Utility-Preserving Coreset Selection for Balanced Unlearning
User specifications or legal frameworks often require information to be removed from pretrained models, including large language models (LLMs). This requires deleting or "forgetting" a set of data points from an already-trained model, which typically degrades its performance on other data points. Thus, a balance must be struck between removing information and keeping the model's other abilities intact, with a failure to balance this trade-off leading to poor deletion or an unusable model. To this end, we propose UPCORE (Utility-Preserving Coreset Selection), a method-agnostic data selection framework for mitigating collateral damage during unlearning. Finding that the model damage is correlated with the variance of the model's representations on the forget set, we selectively prune the forget set to remove outliers, thereby minimizing model degradation after unlearning. Across three standard unlearning methods, UPCORE consistently achieves a superior balance between the competing objectives of deletion efficacy and model preservation. To better evaluate this trade-off, we introduce a new metric, measuring the area-under-the-curve (AUC) across standard metrics. Our results show that UPCORE improves both standard metrics and AUC, benefiting from positive transfer between the coreset and pruned points while reducing negative transfer from the forget set to points outside of it.
comment: Code: https://github.com/Vaidehi99/UPCORE
♻ ☆ A Multi-Stage Framework with Taxonomy-Guided Reasoning for Occupation Classification Using Large Language Models
Automatically annotating job data with standardized occupations from taxonomies, known as occupation classification, is crucial for labor market analysis. However, this task is often hindered by data scarcity and the challenges of manual annotations. While large language models (LLMs) hold promise due to their extensive world knowledge and in-context learning capabilities, their effectiveness depends on their knowledge of occupational taxonomies, which remains unclear. In this study, we assess the ability of LLMs to generate precise taxonomic entities from taxonomy, highlighting their limitations, especially for smaller models. To address these challenges, we propose a multi-stage framework consisting of inference, retrieval, and reranking stages, which integrates taxonomy-guided reasoning examples to enhance performance by aligning outputs with taxonomic knowledge. Evaluations on a large-scale dataset show that our framework not only enhances occupation and skill classification tasks, but also provides a cost-effective alternative to frontier models like GPT-4o, significantly reducing computational costs while maintaining strong performance. This makes it a practical and scalable solution for occupation classification and related tasks across LLMs.
comment: Accepted to ICWSM'26
♻ ☆ Fine-Tune an SLM or Prompt an LLM? The Case of Generating Low-Code Workflows KDD 2025
Large Language Models (LLMs) such as GPT-4o can handle a wide range of complex tasks with the right prompt. As per token costs are reduced, the advantages of fine-tuning Small Language Models (SLMs) for real-world applications -- faster inference, lower costs -- may no longer be clear. In this work, we present evidence that, for domain-specific tasks that require structured outputs, SLMs still have a quality advantage. We compare fine-tuning an SLM against prompting LLMs on the task of generating low-code workflows in JSON form. We observe that while a good prompt can yield reasonable results, fine-tuning improves quality by 10% on average. We also perform systematic error analysis to reveal model limitations.
comment: 8 pages, 7 figures. Accepted to Workshop on Structured Knowledge for Large Language Models (SKnowLLM) at KDD 2025
♻ ☆ Cross-Layer Discrete Concept Discovery for Interpreting Language Models
Uncovering emergent concepts across transformer layers remains a significant challenge because the residual stream linearly mixes and duplicates information, obscuring how features evolve within large language models. Current research efforts primarily inspect neural representations at single layers, thereby overlooking this cross-layer superposition and the redundancy it introduces. These representations are typically either analyzed directly for activation patterns or passed to probing classifiers that map them to a limited set of predefined concepts. To address these limitations, we propose cross-layer VQ-VAE (CLVQ-VAE), a framework that uses vector quantization to map representations across layers and in the process collapse duplicated residual-stream features into compact, interpretable concept vectors. Our approach uniquely combines top-k temperature-based sampling during quantization with EMA codebook updates, providing controlled exploration of the discrete latent space while maintaining code-book diversity. We further enhance the framework with scaled-spherical k-means++ for codebook initialization, which clusters by directional similarity rather than magnitude, better aligning with semantic structure in word embedding space.
♻ ☆ Multi-task retriever fine-tuning for domain-specific and efficient RAG KDD 2025
Retrieval-Augmented Generation (RAG) has become ubiquitous when deploying Large Language Models (LLMs), as it can address typical limitations such as generating hallucinated or outdated information. However, when building real-world RAG applications, practical issues arise. First, the retrieved information is generally domain-specific. Since it is computationally expensive to fine-tune LLMs, it is more feasible to fine-tune the retriever to improve the quality of the data included in the LLM input. Second, as more applications are deployed in the same real-world system, one cannot afford to deploy separate retrievers. Moreover, these RAG applications normally retrieve different kinds of data. Our solution is to instruction fine-tune a small retriever encoder on a variety of domain-specific tasks to allow us to deploy one encoder that can serve many use cases, thereby achieving low-cost, scalability, and speed. We show how this encoder generalizes to out-of-domain settings as well as to an unseen retrieval task on real-world enterprise use cases.
comment: 7 pages, 2 figures. Accepted at Workshop on Structured Knowledge for Large Language Models (SKnowLLM) at KDD 2025
♻ ☆ LoRA Done RITE: Robust Invariant Transformation Equilibration for LoRA Optimization ICLR 2025
Low-rank adaption (LoRA) is a widely used parameter-efficient finetuning method for LLM that reduces memory requirements. However, current LoRA optimizers lack transformation invariance, meaning the actual updates to the weights depends on how the two LoRA factors are scaled or rotated. This deficiency leads to inefficient learning and sub-optimal solutions in practice. This paper introduces LoRA-RITE, a novel adaptive matrix preconditioning method for LoRA optimization, which can achieve transformation invariance and remain computationally efficient. We provide theoretical analysis to demonstrate the benefit of our method and conduct experiments on various LLM tasks with different models including Gemma 2B, 7B, and mT5-XXL. The results demonstrate consistent improvements against existing optimizers. For example, replacing Adam with LoRA-RITE during LoRA fine-tuning of Gemma-2B yielded 4.6\% accuracy gain on Super-Natural Instructions and 3.5\% accuracy gain across other four LLM benchmarks (HellaSwag, ArcChallenge, GSM8K, OpenBookQA).
comment: Published as an oral paper at ICLR 2025. The code for our project is available at https://github.com/gkevinyen5418/LoRA-RITE
♻ ☆ SCULPT: Systematic Tuning of Long Prompts ACL
Prompt optimization is essential for effective utilization of large language models (LLMs) across diverse tasks. While existing optimization methods are effective in optimizing short prompts, they struggle with longer, more complex ones, often risking information loss and being sensitive to small perturbations. To address these challenges, we propose SCULPT (Systematic Tuning of Long Prompts), a framework that treats prompt optimization as a hierarchical tree refinement problem. SCULPT represents prompts as tree structures, enabling targeted modifications while preserving contextual integrity. It employs a Critic-Actor framework that generates reflections and applies actions to refine the prompt. Evaluations demonstrate SCULPT's effectiveness on long prompts, its robustness to adversarial perturbations, and its ability to generate high-performing prompts even without any initial human-written prompt. Compared to existing state of the art methods, SCULPT consistently improves LLM performance by preserving essential task information while applying structured refinements. Both qualitative and quantitative analyses show that SCULPT produces more stable and interpretable prompt modifications, ensuring better generalization across tasks.
comment: Accepted at ACL Main 2025
♻ ☆ Memorization Inheritance in Sequence-Level Knowledge Distillation for Neural Machine Translation ACL 2025
In this work, we explore how instance-level memorization in the teacher Neural Machine Translation (NMT) model gets inherited by the student model in sequence-level knowledge distillation (SeqKD). We find that despite not directly seeing the original training data, students memorize more than baseline models (models of the same size, trained on the original data) -- 3.4% for exact matches and 57% for extractive memorization -- and show increased hallucination rates. Further, under this SeqKD setting, we also characterize how students behave on specific training data subgroups, such as subgroups with low quality and specific counterfactual memorization (CM) scores, and find that students exhibit amplified denoising on low-quality subgroups. Finally, we propose a modification to SeqKD named Adaptive-SeqKD, which intervenes in SeqKD to reduce memorization and hallucinations. Overall, we recommend caution when applying SeqKD: students inherit both their teachers' superior performance and their fault modes, thereby requiring active monitoring.
comment: To appear at ACL 2025; 15 pages total (5 in the main paper, 3 pages of limitations and references and 7 pages with appendices)
♻ ☆ What Factors Affect LLMs and RLLMs in Financial Question Answering?
Recently, the development of large language models (LLMs) and reasoning large language models (RLLMs) have gained considerable attention from many researchers. RLLMs enhance the reasoning capabilities of LLMs through Long Chain-of-Thought (Long CoT) processes, significantly improving the performance of LLMs in addressing complex problems. However, there are few works that systematically explore what methods can fully unlock the performance of LLMs and RLLMs within the financial domain. To investigate the impact of various methods on LLMs and RLLMs, we utilize five LLMs and three RLLMs to assess the effects of prompting methods, agentic frameworks, and multilingual alignment methods on financial question-answering tasks. Our research findings indicate: (1) Current prompting methods and agent frameworks enhance the performance of LLMs in financial question answering by simulating Long CoT; (2) RLLMs possess inherent Long CoT capabilities, which limits the effectiveness of conventional methods in further enhancing their performance; (3) Current advanced multilingual alignment methods primarily improve the multilingual performance of LLMs by extending the reasoning length, which yields minimal benefits for RLLMs. We hope that this study can serve as an important reference for LLMs and RLLMs in the field of financial question answering.
comment: Preprint
♻ ☆ Prompt Perturbations Reveal Human-Like Biases in LLM Survey Responses
Large Language Models (LLMs) are increasingly used as proxies for human subjects in social science surveys, but their reliability and susceptibility to known response biases are poorly understood. This paper investigates the response robustness of LLMs in normative survey contexts - we test nine diverse LLMs on questions from the World Values Survey (WVS), applying a comprehensive set of 11 perturbations to both question phrasing and answer option structure, resulting in over 167,000 simulated interviews. In doing so, we not only reveal LLMs' vulnerabilities to perturbations but also show that all tested models exhibit a consistent recency bias varying in intensity, disproportionately favoring the last-presented answer option. While larger models are generally more robust, all models remain sensitive to semantic variations like paraphrasing and to combined perturbations. By applying a set of perturbations, we reveal that LLMs partially align with survey response biases identified in humans. This underscores the critical importance of prompt design and robustness testing when using LLMs to generate synthetic survey data.
comment: 18 pages, 17 figures
Computer Vision and Pattern Recognition 15
☆ PhysX: Physical-Grounded 3D Asset Generation
3D modeling is moving from virtual to physical. Existing 3D generation primarily emphasizes geometries and textures while neglecting physical-grounded modeling. Consequently, despite the rapid development of 3D generative models, the synthesized 3D assets often overlook rich and important physical properties, hampering their real-world application in physical domains like simulation and embodied AI. As an initial attempt to address this challenge, we propose \textbf{PhysX}, an end-to-end paradigm for physical-grounded 3D asset generation. 1) To bridge the critical gap in physics-annotated 3D datasets, we present PhysXNet - the first physics-grounded 3D dataset systematically annotated across five foundational dimensions: absolute scale, material, affordance, kinematics, and function description. In particular, we devise a scalable human-in-the-loop annotation pipeline based on vision-language models, which enables efficient creation of physics-first assets from raw 3D assets.2) Furthermore, we propose \textbf{PhysXGen}, a feed-forward framework for physics-grounded image-to-3D asset generation, injecting physical knowledge into the pre-trained 3D structural space. Specifically, PhysXGen employs a dual-branch architecture to explicitly model the latent correlations between 3D structures and physical properties, thereby producing 3D assets with plausible physical predictions while preserving the native geometry quality. Extensive experiments validate the superior performance and promising generalization capability of our framework. All the code, data, and models will be released to facilitate future research in generative physical AI.
comment: Project page: https://physx-3d.github.io/
☆ CytoSAE: Interpretable Cell Embeddings for Hematology
Sparse autoencoders (SAEs) emerged as a promising tool for mechanistic interpretability of transformer-based foundation models. Very recently, SAEs were also adopted for the visual domain, enabling the discovery of visual concepts and their patch-wise attribution to tokens in the transformer model. While a growing number of foundation models emerged for medical imaging, tools for explaining their inferences are still lacking. In this work, we show the applicability of SAEs for hematology. We propose CytoSAE, a sparse autoencoder which is trained on over 40,000 peripheral blood single-cell images. CytoSAE generalizes to diverse and out-of-domain datasets, including bone marrow cytology, where it identifies morphologically relevant concepts which we validated with medical experts. Furthermore, we demonstrate scenarios in which CytoSAE can generate patient-specific and disease-specific concepts, enabling the detection of pathognomonic cells and localized cellular abnormalities at the patch level. We quantified the effect of concepts on a patient-level AML subtype classification task and show that CytoSAE concepts reach performance comparable to the state-of-the-art, while offering explainability on the sub-cellular level. Source code and model weights are available at https://github.com/dynamical-inference/cytosae.
comment: 11 pages, 5 figures
☆ MMHU: A Massive-Scale Multimodal Benchmark for Human Behavior Understanding
Humans are integral components of the transportation ecosystem, and understanding their behaviors is crucial to facilitating the development of safe driving systems. Although recent progress has explored various aspects of human behavior$\unicode{x2014}$such as motion, trajectories, and intention$\unicode{x2014}$a comprehensive benchmark for evaluating human behavior understanding in autonomous driving remains unavailable. In this work, we propose $\textbf{MMHU}$, a large-scale benchmark for human behavior analysis featuring rich annotations, such as human motion and trajectories, text description for human motions, human intention, and critical behavior labels relevant to driving safety. Our dataset encompasses 57k human motion clips and 1.73M frames gathered from diverse sources, including established driving datasets such as Waymo, in-the-wild videos from YouTube, and self-collected data. A human-in-the-loop annotation pipeline is developed to generate rich behavior captions. We provide a thorough dataset analysis and benchmark multiple tasks$\unicode{x2014}$ranging from motion prediction to motion generation and human behavior question answering$\unicode{x2014}$thereby offering a broad evaluation suite. Project page : https://MMHU-Benchmark.github.io.
☆ SpatialTrackerV2: 3D Point Tracking Made Easy ICCV 2025
We present SpatialTrackerV2, a feed-forward 3D point tracking method for monocular videos. Going beyond modular pipelines built on off-the-shelf components for 3D tracking, our approach unifies the intrinsic connections between point tracking, monocular depth, and camera pose estimation into a high-performing and feedforward 3D point tracker. It decomposes world-space 3D motion into scene geometry, camera ego-motion, and pixel-wise object motion, with a fully differentiable and end-to-end architecture, allowing scalable training across a wide range of datasets, including synthetic sequences, posed RGB-D videos, and unlabeled in-the-wild footage. By learning geometry and motion jointly from such heterogeneous data, SpatialTrackerV2 outperforms existing 3D tracking methods by 30%, and matches the accuracy of leading dynamic 3D reconstruction approaches while running 50$\times$ faster.
comment: International Conference on Computer Vision, ICCV 2025. Huggingface Demo: https://huggingface.co/spaces/Yuxihenry/SpatialTrackerV2, Code: https://github.com/henry123-boy/SpaTrackerV2
☆ Interpreting Radiologist's Intention from Eye Movements in Chest X-ray Diagnosis ACM MM 2025
Radiologists rely on eye movements to navigate and interpret medical images. A trained radiologist possesses knowledge about the potential diseases that may be present in the images and, when searching, follows a mental checklist to locate them using their gaze. This is a key observation, yet existing models fail to capture the underlying intent behind each fixation. In this paper, we introduce a deep learning-based approach, RadGazeIntent, designed to model this behavior: having an intention to find something and actively searching for it. Our transformer-based architecture processes both the temporal and spatial dimensions of gaze data, transforming fine-grained fixation features into coarse, meaningful representations of diagnostic intent to interpret radiologists' goals. To capture the nuances of radiologists' varied intention-driven behaviors, we process existing medical eye-tracking datasets to create three intention-labeled subsets: RadSeq (Systematic Sequential Search), RadExplore (Uncertainty-driven Exploration), and RadHybrid (Hybrid Pattern). Experimental results demonstrate RadGazeIntent's ability to predict which findings radiologists are examining at specific moments, outperforming baseline methods across all intention-labeled datasets.
comment: ACM MM 2025
☆ Mitigating Object Hallucinations via Sentence-Level Early Intervention
Multimodal large language models (MLLMs) have revolutionized cross-modal understanding but continue to struggle with hallucinations - fabricated content contradicting visual inputs. Existing hallucination mitigation methods either incur prohibitive computational costs or introduce distribution mismatches between training data and model outputs. We identify a critical insight: hallucinations predominantly emerge at the early stages of text generation and propagate through subsequent outputs. To address this, we propose **SENTINEL** (**S**entence-level **E**arly i**N**tervention **T**hrough **IN**-domain pr**E**ference **L**earning), a framework that eliminates dependency on human annotations. Specifically, we first bootstrap high-quality in-domain preference pairs by iteratively sampling model outputs, validating object existence through cross-checking with two open-vocabulary detectors, and classifying sentences into hallucinated/non-hallucinated categories. Subsequently, we use context-coherent positive samples and hallucinated negative samples to build context-aware preference data iteratively. Finally, we train models using a context-aware preference loss (C-DPO) that emphasizes discriminative learning at the sentence level where hallucinations initially manifest. Experimental results show that SENTINEL can reduce hallucinations by over 90\% compared to the original model and outperforms the previous state-of-the-art method on both hallucination benchmarks and general capabilities benchmarks, demonstrating its superiority and generalization ability. The models, datasets, and code are available at https://github.com/pspdada/SENTINEL.
☆ Vision-based Perception for Autonomous Vehicles in Obstacle Avoidance Scenarios
Obstacle avoidance is essential for ensuring the safety of autonomous vehicles. Accurate perception and motion planning are crucial to enabling vehicles to navigate complex environments while avoiding collisions. In this paper, we propose an efficient obstacle avoidance pipeline that leverages a camera-only perception module and a Frenet-Pure Pursuit-based planning strategy. By integrating advancements in computer vision, the system utilizes YOLOv11 for object detection and state-of-the-art monocular depth estimation models, such as Depth Anything V2, to estimate object distances. A comparative analysis of these models provides valuable insights into their accuracy, efficiency, and robustness in real-world conditions. The system is evaluated in diverse scenarios on a university campus, demonstrating its effectiveness in handling various obstacles and enhancing autonomous navigation. The video presenting the results of the obstacle avoidance experiments is available at: https://www.youtube.com/watch?v=FoXiO5S_tA8
comment: 7 pages, 6 figures, 4 tables, HSI 2025
☆ Describe Anything Model for Visual Question Answering on Text-rich Images ICCV 2025
Recent progress has been made in region-aware vision-language modeling, particularly with the emergence of the Describe Anything Model (DAM). DAM is capable of generating detailed descriptions of any specific image areas or objects without the need for additional localized image-text alignment supervision. We hypothesize that such region-level descriptive capability is beneficial for the task of Visual Question Answering (VQA), especially in challenging scenarios involving images with dense text. In such settings, the fine-grained extraction of textual information is crucial to producing correct answers. Motivated by this, we introduce DAM-QA, a framework with a tailored evaluation protocol, developed to investigate and harness the region-aware capabilities from DAM for the text-rich VQA problem that requires reasoning over text-based information within images. DAM-QA incorporates a mechanism that aggregates answers from multiple regional views of image content, enabling more effective identification of evidence that may be tied to text-related elements. Experiments on six VQA benchmarks show that our approach consistently outperforms the baseline DAM, with a notable 7+ point gain on DocVQA. DAM-QA also achieves the best overall performance among region-aware models with fewer parameters, significantly narrowing the gap with strong generalist VLMs. These results highlight the potential of DAM-like models for text-rich and broader VQA tasks when paired with efficient usage and integration strategies. Our code is publicly available at https://github.com/Linvyl/DAM-QA.git.
comment: 11 pages, 5 figures. Accepted to VisionDocs @ ICCV 2025
☆ EgoVLA: Learning Vision-Language-Action Models from Egocentric Human Videos
Real robot data collection for imitation learning has led to significant advancements in robotic manipulation. However, the requirement for robot hardware in the process fundamentally constrains the scale of the data. In this paper, we explore training Vision-Language-Action (VLA) models using egocentric human videos. The benefit of using human videos is not only for their scale but more importantly for the richness of scenes and tasks. With a VLA trained on human video that predicts human wrist and hand actions, we can perform Inverse Kinematics and retargeting to convert the human actions to robot actions. We fine-tune the model using a few robot manipulation demonstrations to obtain the robot policy, namely EgoVLA. We propose a simulation benchmark called Isaac Humanoid Manipulation Benchmark, where we design diverse bimanual manipulation tasks with demonstrations. We fine-tune and evaluate EgoVLA with Isaac Humanoid Manipulation Benchmark and show significant improvements over baselines and ablate the importance of human data. Videos can be found on our website: https://rchalyang.github.io/EgoVLA
comment: More videos can be found on our website: https://rchalyang.github.io/EgoVLA
☆ Traffic-Aware Pedestrian Intention Prediction
Accurate pedestrian intention estimation is crucial for the safe navigation of autonomous vehicles (AVs) and hence attracts a lot of research attention. However, current models often fail to adequately consider dynamic traffic signals and contextual scene information, which are critical for real-world applications. This paper presents a Traffic-Aware Spatio-Temporal Graph Convolutional Network (TA-STGCN) that integrates traffic signs and their states (Red, Yellow, Green) into pedestrian intention prediction. Our approach introduces the integration of dynamic traffic signal states and bounding box size as key features, allowing the model to capture both spatial and temporal dependencies in complex urban environments. The model surpasses existing methods in accuracy. Specifically, TA-STGCN achieves a 4.75% higher accuracy compared to the baseline model on the PIE dataset, demonstrating its effectiveness in improving pedestrian intention prediction.
comment: 6 pages, 4 figures. Accepted to the American Control Conference (ACC) 2025
☆ Unit-Based Histopathology Tissue Segmentation via Multi-Level Feature Representation
We propose UTS, a unit-based tissue segmentation framework for histopathology that classifies each fixed-size 32 * 32 tile, rather than each pixel, as the segmentation unit. This approach reduces annotation effort and improves computational efficiency without compromising accuracy. To implement this approach, we introduce a Multi-Level Vision Transformer (L-ViT), which benefits the multi-level feature representation to capture both fine-grained morphology and global tissue context. Trained to segment breast tissue into three categories (infiltrating tumor, non-neoplastic stroma, and fat), UTS supports clinically relevant tasks such as tumor-stroma quantification and surgical margin assessment. Evaluated on 386,371 tiles from 459 H&E-stained regions, it outperforms U-Net variants and transformer-based baselines. Code and Dataset will be available at GitHub.
comment: 12 pages, 6 figures
♻ ☆ MAMBO: High-Resolution Generative Approach for Mammography Images
Mammography is the gold standard for the detection and diagnosis of breast cancer. This procedure can be significantly enhanced with Artificial Intelligence (AI)-based software, which assists radiologists in identifying abnormalities. However, training AI systems requires large and diverse datasets, which are often difficult to obtain due to privacy and ethical constraints. To address this issue, the paper introduces MAMmography ensemBle mOdel (MAMBO), a novel patch-based diffusion approach designed to generate full-resolution mammograms. Diffusion models have shown breakthrough results in realistic image generation, yet few studies have focused on mammograms, and none have successfully generated high-resolution outputs required to capture fine-grained features of small lesions. To achieve this, MAMBO integrates separate diffusion models to capture both local and global (image-level) contexts. The contextual information is then fed into the final model, significantly aiding the noise removal process. This design enables MAMBO to generate highly realistic mammograms of up to 3840x3840 pixels. Importantly, this approach can be used to enhance the training of classification models and extended to anomaly segmentation. Experiments, both numerical and radiologist validation, assess MAMBO's capabilities in image generation, super-resolution, and anomaly segmentation, highlighting its potential to enhance mammography analysis for more accurate diagnoses and earlier lesion detection. The source code used in this study is publicly available at: https://github.com/iai-rs/mambo.
comment: 21 pages, 14 figures, 7 tables This work has been submitted to the IEEE for possible publication
♻ ☆ Flow-GRPO: Training Flow Matching Models via Online RL
We propose Flow-GRPO, the first method integrating online reinforcement learning (RL) into flow matching models. Our approach uses two key strategies: (1) an ODE-to-SDE conversion that transforms a deterministic Ordinary Differential Equation (ODE) into an equivalent Stochastic Differential Equation (SDE) that matches the original model's marginal distribution at all timesteps, enabling statistical sampling for RL exploration; and (2) a Denoising Reduction strategy that reduces training denoising steps while retaining the original inference timestep number, significantly improving sampling efficiency without performance degradation. Empirically, Flow-GRPO is effective across multiple text-to-image tasks. For complex compositions, RL-tuned SD3.5 generates nearly perfect object counts, spatial relations, and fine-grained attributes, boosting GenEval accuracy from 63% to 95%. In visual text rendering, its accuracy improves from 59% to 92%, significantly enhancing text generation. Flow-GRPO also achieves substantial gains in human preference alignment. Notably, very little reward hacking occurred, meaning rewards did not increase at the cost of appreciable image quality or diversity degradation.
comment: Code: https://github.com/yifan123/flow_grpo
♻ ☆ 4D-MISR: A unified model for low-dose super-resolution imaging via feature fusion
While electron microscopy offers crucial atomic-resolution insights into structure-property relationships, radiation damage severely limits its use on beam-sensitive materials like proteins and 2D materials. To overcome this challenge, we push beyond the electron dose limits of conventional electron microscopy by adapting principles from multi-image super-resolution (MISR) that have been widely used in remote sensing. Our method fuses multiple low-resolution, sub-pixel-shifted views and enhances the reconstruction with a convolutional neural network (CNN) that integrates features from synthetic, multi-angle observations. We developed a dual-path, attention-guided network for 4D-STEM that achieves atomic-scale super-resolution from ultra-low-dose data. This provides robust atomic-scale visualization across amorphous, semi-crystalline, and crystalline beam-sensitive specimens. Systematic evaluations on representative materials demonstrate comparable spatial resolution to conventional ptychography under ultra-low-dose conditions. Our work expands the capabilities of 4D-STEM, offering a new and generalizable method for the structural analysis of radiation-vulnerable materials.
comment: The authors have decided to withdraw this submission due to incorrect affiliation information and certain logical inconsistencies in the manuscript that require revision
♻ ☆ DUNIA: Pixel-Sized Embeddings via Cross-Modal Alignment for Earth Observation Applications
Significant efforts have been directed towards adapting self-supervised multimodal learning for Earth observation applications. However, most current methods produce coarse patch-sized embeddings, limiting their effectiveness and integration with other modalities like LiDAR. To close this gap, we present DUNIA, an approach to learn pixel-sized embeddings through cross-modal alignment between images and full-waveform LiDAR data. As the model is trained in a contrastive manner, the embeddings can be directly leveraged in the context of a variety of environmental monitoring tasks in a zero-shot setting. In our experiments, we demonstrate the effectiveness of the embeddings for seven such tasks: canopy height mapping, fractional canopy cover, land cover mapping, tree species identification, plant area index, crop type classification, and per-pixel waveform-based vertical structure mapping. The results show that the embeddings, along with zero-shot classifiers, often outperform specialized supervised models, even in low-data regimes. In the fine-tuning setting, we show strong performances near or better than the state-of-the-art on five out of six tasks.
comment: 26 pages, 8 figures
Information Retrieval 15
☆ Advancing Retrieval-Augmented Generation for Structured Enterprise and Internal Data
Organizations increasingly rely on proprietary enterprise data, including HR records, structured reports, and tabular documents, for critical decision-making. While Large Language Models (LLMs) have strong generative capabilities, they are limited by static pretraining, short context windows, and challenges in processing heterogeneous data formats. Conventional Retrieval-Augmented Generation (RAG) frameworks address some of these gaps but often struggle with structured and semi-structured data. This work proposes an advanced RAG framework that combines hybrid retrieval strategies using dense embeddings (all-mpnet-base-v2) and BM25, enhanced by metadata-aware filtering with SpaCy NER and cross-encoder reranking. The framework applies semantic chunking to maintain textual coherence and retains tabular data structures to preserve row-column integrity. Quantized indexing optimizes retrieval efficiency, while human-in-the-loop feedback and conversation memory improve adaptability. Experiments on enterprise datasets show notable improvements: Precision@5 increased by 15 percent (90 versus 75), Recall@5 by 13 percent (87 versus 74), and Mean Reciprocal Rank by 16 percent (0.85 versus 0.69). Qualitative evaluations show higher scores in Faithfulness (4.6 versus 3.0), Completeness (4.2 versus 2.5), and Relevance (4.5 versus 3.2) on a 5-point Likert scale. These results demonstrate the framework's effectiveness in delivering accurate, comprehensive, and contextually relevant responses for enterprise tasks. Future work includes extending to multimodal data and integrating agent-based retrieval. The source code will be released at https://github.com/CheerlaChandana/Enterprise-Chatbot
☆ Developing Visual Augmented Q&A System using Scalable Vision Embedding Retrieval & Late Interaction Re-ranker SIGIR
Traditional information extraction systems face challenges with text only language models as it does not consider infographics (visual elements of information) such as tables, charts, images etc. often used to convey complex information to readers. Multimodal LLM (MLLM) face challenges of finding needle in the haystack problem i.e., either longer context length or substantial number of documents as search space. Late interaction mechanism over visual language models has shown state of the art performance in retrieval-based vision augmented Q&A tasks. There are yet few challenges using it for RAG based multi-modal Q&A. Firstly, many popular and widely adopted vector databases do not support native multi-vector retrieval. Secondly, late interaction requires computation which inflates space footprint and can hinder enterprise adoption. Lastly, the current state of late interaction mechanism does not leverage the approximate neighbor search indexing methods for large speed ups in retrieval process. This paper explores a pragmatic approach to make vision retrieval process scalable and efficient without compromising on performance quality. We propose multi-step custom implementation utilizing widely adopted hybrid search (metadata & embedding) and state of the art late interaction re-ranker to retrieve best matching pages. Finally, MLLM are prompted as reader to generate answers from contextualized best matching pages. Through experiments, we observe that the proposed design is scalable (significant speed up) and stable (without degrading performance quality), hence can be used as production systems at enterprises.
comment: Presented at NLP@IR workshop at SIGIR conference
☆ An Ecosystem for Ontology Interoperability
Ontology interoperability is one of the complicated issues that restricts the use of ontologies in knowledge graphs (KGs). Different ontologies with conflicting and overlapping concepts make it difficult to design, develop, and deploy an interoperable ontology for downstream tasks. We propose an ecosystem for ontology interoperability. The ecosystem employs three state-of-the-art semantic techniques in different phases of the ontology engineering life cycle: ontology design patterns (ODPs) in the design phase, ontology matching and versioning (OM\&OV) in the develop phase, and ontology-compliant knowledge graphs (OCKGs) in the deploy phase, to achieve better ontology interoperability in real-world applications. A case study in the building domain validates the usefulness of the proposed ecosystem.
comment: 4 pages, 8 figures
☆ Looking for Fairness in Recommender Systems
Recommender systems can be found everywhere today, shaping our everyday experience whenever we're consuming content, ordering food, buying groceries online, or even just reading the news. Let's imagine we're in the process of building a recommender system to make content suggestions to users on social media. When thinking about fairness, it becomes clear there are several perspectives to consider: the users asking for tailored suggestions, the content creators hoping for some limelight, and society at large, navigating the repercussions of algorithmic recommendations. A shared fairness concern across all three is the emergence of filter bubbles, a side-effect that takes place when recommender systems are almost "too good", making recommendations so tailored that users become inadvertently confined to a narrow set of opinions/themes and isolated from alternative ideas. From the user's perspective, this is akin to manipulation. From the small content creator's perspective, this is an obstacle preventing them access to a whole range of potential fans. From society's perspective, the potential consequences are far-reaching, influencing collective opinions, social behavior and political decisions. How can our recommender system be fine-tuned to avoid the creation of filter bubbles, and ensure a more inclusive and diverse content landscape? Approaching this problem involves defining one (or more) performance metric to represent diversity, and tweaking our recommender system's performance through the lens of fairness. By incorporating this metric into our evaluation framework, we aim to strike a balance between personalized recommendations and the broader societal goal of fostering rich and varied cultures and points of view.
☆ Sparse Autoencoders for Sequential Recommendation Models: Interpretation and Flexible Control
Many current state-of-the-art models for sequential recommendations are based on transformer architectures. Interpretation and explanation of such black box models is an important research question, as a better understanding of their internals can help understand, influence, and control their behavior, which is very important in a variety of real-world applications. Recently sparse autoencoders (SAE) have been shown to be a promising unsupervised approach for extracting interpretable features from language models. These autoencoders learn to reconstruct hidden states of the transformer's internal layers from sparse linear combinations of directions in their activation space. This paper is focused on the application of SAE to the sequential recommendation domain. We show that this approach can be successfully applied to the transformer trained on a sequential recommendation task: learned directions turn out to be more interpretable and monosemantic than the original hidden state dimensions. Moreover, we demonstrate that the features learned by SAE can be used to effectively and flexibly control the model's behavior, providing end-users with a straightforward method to adjust their recommendations to different custom scenarios and contexts.
☆ AFPM: Alignment-based Frame Patch Modeling for Cross-Dataset EEG Decoding
Electroencephalogram (EEG) decoding models for brain-computer interfaces (BCIs) struggle with cross-dataset learning and generalization due to channel layout inconsistencies, non-stationary signal distributions, and limited neurophysiological prior integration. To address these issues, we propose a plug-and-play Alignment-Based Frame-Patch Modeling (AFPM) framework, which has two main components: 1) Spatial Alignment, which selects task-relevant channels based on brain-region priors, aligns EEG distributions across domains, and remaps the selected channels to a unified layout; and, 2) Frame-Patch Encoding, which models multi-dataset signals into unified spatiotemporal patches for EEG decoding. Compared to 17 state-of-the-art approaches that need dataset-specific tuning, the proposed calibration-free AFPM achieves performance gains of up to 4.40% on motor imagery and 3.58% on event-related potential tasks. To our knowledge, this is the first calibration-free cross-dataset EEG decoding framework, substantially enhancing the practicalness of BCIs in real-world applications.
☆ SIEVE: Effective Filtered Vector Search with Collection of Indexes
Many real-world tasks such as recommending videos with the kids tag can be reduced to finding most similar vectors associated with hard predicates. This task, filtered vector search, is challenging as prior state-of-the-art graph-based (unfiltered) similarity search techniques quickly degenerate when hard constraints are considered. That is, effective graph-based filtered similarity search relies on sufficient connectivity for reaching the most similar items within just a few hops. To consider predicates, recent works propose modifying graph traversal to visit only the items that may satisfy predicates. However, they fail to offer the just-a-few-hops property for a wide range of predicates: they must restrict predicates significantly or lose efficiency if only a small fraction of items satisfy predicates. We propose an opposite approach: instead of constraining traversal, we build many indexes each serving different predicate forms. For effective construction, we devise a three-dimensional analytical model capturing relationships among index size, search time, and recall, with which we follow a workload-aware approach to pack as many useful indexes as possible into a collection. At query time, the analytical model is employed yet again to discern the one that offers the fastest search at a given recall. We show superior performance and support on datasets with varying selectivities and forms: our approach achieves up to 8.06x speedup while having as low as 1% build time versus other indexes, with less than 2.15x memory of a standard HNSW graph and modest knowledge of past workloads.
☆ Context-Aware Search and Retrieval Over Erasure Channels
This paper introduces and analyzes a search and retrieval model that adopts key semantic communication principles from retrieval-augmented generation. We specifically present an information-theoretic analysis of a remote document retrieval system operating over a symbol erasure channel. The proposed model encodes the feature vector of a query, derived from term-frequency weights of a language corpus by using a repetition code with an adaptive rate dependent on the contextual importance of the terms. At the decoder, we select between two documents based on the contextual closeness of the recovered query. By leveraging a jointly Gaussian approximation for both the true and reconstructed similarity scores, we derive an explicit expression for the retrieval error probability, i.e., the probability under which the less similar document is selected. Numerical simulations on synthetic and real-world data (Google NQ) confirm the validity of the analysis. They further demonstrate that assigning greater redundancy to critical features effectively reduces the error rate, highlighting the effectiveness of semantic-aware feature encoding in error-prone communication settings.
☆ Similarity-Guided Diffusion for Contrastive Sequential Recommendation
In sequential recommendation systems, data augmentation and contrastive learning techniques have recently been introduced using diffusion models to achieve robust representation learning. However, most of the existing approaches use random augmentation, which risk damaging the contextual information of the original sequence. Accordingly, we propose a Similarity-Guided Diffusion for Contrastive Sequential Recommendation. Our method leverages the similarity between item embedding vectors to generate semantically consistent noise. Moreover, we utilize high confidence score in the denoising process to select our augmentation positions. This approach more effectively reflects contextual and structural information compared to augmentation at random positions. From a contrastive learning perspective, the proposed augmentation technique provides more discriminative positive and negative samples, simultaneously improving training efficiency and recommendation performance. Experimental results on five benchmark datasets show that SimDiffRec outperforms the existing baseline models.
comment: 14 pages, 5 figures
♻ ☆ Protecting Copyrighted Material with Unique Identifiers in Large Language Model Training
A primary concern regarding training large language models (LLMs) is whether they abuse copyrighted online text. With the increasing training data scale and the prevalence of LLMs in daily lives, two problems arise: \textbf{1)} false positive membership inference results misled by similar examples; \textbf{2)} membership inference methods are usually too complex for end users to understand and use. To address these issues, we propose an alternative \textit{insert-and-detect} methodology, advocating that web users and content platforms employ \textbf{\textit{unique identifiers}} for reliable and independent membership inference. Users and platforms can create their identifiers, embed them in copyrighted text, and independently detect them in future LLMs. As an initial demonstration, we introduce \textit{\textbf{ghost sentences}} and a user-friendly last-$k$ words test, allowing end users to chat with LLMs for membership inference. Ghost sentences consist primarily of unique passphrases of random natural words, which can come with customized elements to bypass possible filter rules. The last-$k$ words test requires a significant repetition time of ghost sentences~($\ge10$). For cases with fewer repetitions, we designed an extra perplexity test, as LLMs exhibit high perplexity when encountering unnatural passphrases. We also conduct a comprehensive study on the memorization and membership inference of ghost sentences, examining factors such as training data scales, model sizes, repetition times, insertion positions, wordlist of passphrases, alignment, \textit{etc}. Our study shows the possibility of applying ghost sentences in real scenarios and provides instructions for the potential application.
comment: A technical report, work mainly done in the early of 2024
♻ ☆ A Multistakeholder Approach to Value-Driven Co-Design of Recommender System Evaluation Metrics in Digital Archives RecSys 2025
This paper presents the first multistakeholder approach for translating diverse stakeholder values into an evaluation metric setup for Recommender Systems (RecSys) in digital archives. While commercial platforms mainly rely on engagement metrics, cultural heritage domains require frameworks that balance competing priorities among archivists, platform owners, researchers, and other stakeholders. To address this challenge, we conducted high-profile focus groups (5 groups x 5 persons) with upstream, provider, system, consumer, and downstream stakeholders, identifying value priorities across critical dimensions: visibility/representation, expertise adaptation, and transparency/trust. Our analysis shows that stakeholder concerns naturally align with four sequential research funnel stages: discovery, interaction, integration, and impact. The resulting evaluation setup addresses domain-specific challenges including collection representation imbalances, non-linear research patterns, and tensions between specialized expertise and broader accessibility. We propose directions for tailored metrics in each stage of this research journey, such as research path quality for discovery, contextual appropriateness for interaction, metadata-weighted relevance for integration, and cross-stakeholder value alignment for impact assessment. Our contributions extend beyond digital archives to the broader RecSys community, offering transferable evaluation approaches for domains where value emerges through sustained engagement rather than immediate consumption.
comment: Accepted at RecSys 2025
♻ ☆ LEADRE: Multi-Faceted Knowledge Enhanced LLM Empowered Display Advertisement Recommender System VLDB 2025
Display advertising provides significant value to advertisers, publishers, and users. Traditional display advertising systems utilize a multi-stage architecture consisting of retrieval, coarse ranking, and final ranking. However, conventional retrieval methods rely on ID-based learning to rank mechanisms and fail to adequately utilize the content information of ads, which hampers their ability to provide diverse recommendation lists. To address this limitation, we propose leveraging the extensive world knowledge of LLMs. However, three key challenges arise when attempting to maximize the effectiveness of LLMs: "How to capture user interests", "How to bridge the knowledge gap between LLMs and advertising system", and "How to efficiently deploy LLMs". To overcome these challenges, we introduce a novel LLM-based framework called LLM Empowered Display ADvertisement REcommender system (LEADRE). LEADRE consists of three core modules: (1) The Intent-Aware Prompt Engineering introduces multi-faceted knowledge and designs intent-aware pairs that fine-tune LLMs to generate ads tailored to users' personal interests. (2) The Advertising-Specific Knowledge Alignment incorporates auxiliary fine-tuning tasks and Direct Preference Optimization (DPO) to align LLMs with ad semantic and business value. (3) The Efficient System Deployment deploys LEADRE in an online environment by integrating both latency-tolerant and latency-sensitive service. Extensive offline experiments demonstrate the effectiveness of LEADRE and validate the contributions of individual modules. Online A/B test shows that LEADRE leads to a 1.57% and 1.17% GMV lift for serviced users on WeChat Channels and Moments separately. LEADRE has been deployed on both platforms, serving tens of billions of requests each day.
comment: Accepted by VLDB 2025 Industrial Track
♻ ☆ Learning an Effective Premise Retrieval Model for Efficient Mathematical Formalization
Formalized mathematics has recently garnered significant attention for its ability to assist mathematicians across various fields. Premise retrieval, as a common step in mathematical formalization, has been a challenge, particularly for inexperienced users. Existing retrieval methods that facilitate natural language queries require a certain level of mathematical expertise from users, while approaches based on formal languages (e.g., Lean) typically struggle with the scarcity of training data, hindering the training of effective and generalizable retrieval models. In this work, we introduce a novel method that leverages data extracted from Mathlib to train a lightweight and effective premise retrieval model. In particular, the proposed model embeds queries (i.e., proof state provided by Lean) and premises in a latent space, featuring a tokenizer specifically trained on formal corpora. The model is learned in a contrastive learning framework, in which a fine-grained similarity calculation method and a re-ranking module are applied to enhance the retrieval performance. Experimental results demonstrate that our model outperforms existing baselines, achieving higher accuracy while maintaining a lower computational load. We have released an open-source search engine based on our retrieval model at https://premise-search.com/. The source code and the trained model can be found at https://github.com/ruc-ai4math/Premise-Retrieval.
♻ ☆ METIS: Fast Quality-Aware RAG Systems with Configuration Adaptation
RAG (Retrieval Augmented Generation) allows LLMs (large language models) to generate better responses with external knowledge, but using more external knowledge often improves generation quality at the expense of response delay. Prior work either reduces the response delay (through better scheduling of RAG queries) or strives to maximize quality (which involves tuning the RAG workflow), but they fall short in optimizing the tradeoff between the delay and quality of RAG responses. This paper presents METIS, the first RAG system that jointly schedules queries and adapts the key RAG configurations of each query, such as the number of retrieved text chunks and synthesis methods, in order to balance quality optimization and response delay reduction. Using 4 popular RAG-QA datasets, we show that compared with the state-of-the-art RAG optimization schemes, METIS reduces the generation latency by $1.64-2.54\times$ without sacrificing generation quality.
comment: 17 pages, 18 figures
♻ ☆ Multi-task retriever fine-tuning for domain-specific and efficient RAG KDD 2025
Retrieval-Augmented Generation (RAG) has become ubiquitous when deploying Large Language Models (LLMs), as it can address typical limitations such as generating hallucinated or outdated information. However, when building real-world RAG applications, practical issues arise. First, the retrieved information is generally domain-specific. Since it is computationally expensive to fine-tune LLMs, it is more feasible to fine-tune the retriever to improve the quality of the data included in the LLM input. Second, as more applications are deployed in the same real-world system, one cannot afford to deploy separate retrievers. Moreover, these RAG applications normally retrieve different kinds of data. Our solution is to instruction fine-tune a small retriever encoder on a variety of domain-specific tasks to allow us to deploy one encoder that can serve many use cases, thereby achieving low-cost, scalability, and speed. We show how this encoder generalizes to out-of-domain settings as well as to an unseen retrieval task on real-world enterprise use cases.
comment: 7 pages, 2 figures. Accepted at Workshop on Structured Knowledge for Large Language Models (SKnowLLM) at KDD 2025
Machine Learning 150
☆ Language Models Improve When Pretraining Data Matches Target Tasks
Every data selection method inherently has a target. In practice, these targets often emerge implicitly through benchmark-driven iteration: researchers develop selection strategies, train models, measure benchmark performance, then refine accordingly. This raises a natural question: what happens when we make this optimization explicit? To explore this, we propose benchmark-targeted ranking (BETR), a simple method that selects pretraining documents based on similarity to benchmark training examples. BETR embeds benchmark examples and a sample of pretraining documents in a shared space, scores this sample by similarity to benchmarks, then trains a lightweight classifier to predict these scores for the full corpus. We compare data selection methods by training over 500 models spanning $10^{19}$ to $10^{22}$ FLOPs and fitting scaling laws to them. From this, we find that simply aligning pretraining data to evaluation benchmarks using BETR achieves a 2.1x compute multiplier over DCLM-Baseline (4.7x over unfiltered data) and improves performance on 9 out of 10 tasks across all scales. BETR also generalizes well: when targeting a diverse set of benchmarks disjoint from our evaluation suite, it still matches or outperforms baselines. Our scaling analysis further reveals a clear trend: larger models require less aggressive filtering. Overall, our findings show that directly matching pretraining data to target tasks precisely shapes model capabilities and highlight that optimal selection strategies must adapt to model scale.
comment: 44 pages, 25 figures, 13 tables
☆ CytoSAE: Interpretable Cell Embeddings for Hematology
Sparse autoencoders (SAEs) emerged as a promising tool for mechanistic interpretability of transformer-based foundation models. Very recently, SAEs were also adopted for the visual domain, enabling the discovery of visual concepts and their patch-wise attribution to tokens in the transformer model. While a growing number of foundation models emerged for medical imaging, tools for explaining their inferences are still lacking. In this work, we show the applicability of SAEs for hematology. We propose CytoSAE, a sparse autoencoder which is trained on over 40,000 peripheral blood single-cell images. CytoSAE generalizes to diverse and out-of-domain datasets, including bone marrow cytology, where it identifies morphologically relevant concepts which we validated with medical experts. Furthermore, we demonstrate scenarios in which CytoSAE can generate patient-specific and disease-specific concepts, enabling the detection of pathognomonic cells and localized cellular abnormalities at the patch level. We quantified the effect of concepts on a patient-level AML subtype classification task and show that CytoSAE concepts reach performance comparable to the state-of-the-art, while offering explainability on the sub-cellular level. Source code and model weights are available at https://github.com/dynamical-inference/cytosae.
comment: 11 pages, 5 figures
☆ Cost-aware Stopping for Bayesian Optimization
In automated machine learning, scientific discovery, and other applications of Bayesian optimization, deciding when to stop evaluating expensive black-box functions is an important practical consideration. While several adaptive stopping rules have been proposed, in the cost-aware setting they lack guarantees ensuring they stop before incurring excessive function evaluation costs. We propose a cost-aware stopping rule for Bayesian optimization that adapts to varying evaluation costs and is free of heuristic tuning. Our rule is grounded in a theoretical connection to state-of-the-art cost-aware acquisition functions, namely the Pandora's Box Gittins Index (PBGI) and log expected improvement per cost. We prove a theoretical guarantee bounding the expected cumulative evaluation cost incurred by our stopping rule when paired with these two acquisition functions. In experiments on synthetic and empirical tasks, including hyperparameter optimization and neural architecture size search, we show that combining our stopping rule with the PBGI acquisition function consistently matches or outperforms other acquisition-function--stopping-rule pairs in terms of cost-adjusted simple regret, a metric capturing trade-offs between solution quality and cumulative evaluation cost.
☆ S2WTM: Spherical Sliced-Wasserstein Autoencoder for Topic Modeling ACL 2025
Modeling latent representations in a hyperspherical space has proven effective for capturing directional similarities in high-dimensional text data, benefiting topic modeling. Variational autoencoder-based neural topic models (VAE-NTMs) commonly adopt the von Mises-Fisher prior to encode hyperspherical structure. However, VAE-NTMs often suffer from posterior collapse, where the KL divergence term in the objective function highly diminishes, leading to ineffective latent representations. To mitigate this issue while modeling hyperspherical structure in the latent space, we propose the Spherical Sliced Wasserstein Autoencoder for Topic Modeling (S2WTM). S2WTM employs a prior distribution supported on the unit hypersphere and leverages the Spherical Sliced-Wasserstein distance to align the aggregated posterior distribution with the prior. Experimental results demonstrate that S2WTM outperforms state-of-the-art topic models, generating more coherent and diverse topics while improving performance on downstream tasks.
comment: Accepted as a long paper for ACL 2025 main conference
☆ Characterizing State Space Model (SSM) and SSM-Transformer Hybrid Language Model Performance with Long Context Length
The demand for machine intelligence capable of processing continuous, long-context inputs on local devices is growing rapidly. However, the quadratic complexity and memory requirements of traditional Transformer architectures make them inefficient and often unusable for these tasks. This has spurred a paradigm shift towards new architectures like State Space Models (SSMs) and hybrids, which promise near-linear scaling. While most current research focuses on the accuracy and theoretical throughput of these models, a systematic performance characterization on practical consumer hardware is critically needed to guide system-level optimization and unlock new applications. To address this gap, we present a comprehensive, comparative benchmarking of carefully selected Transformer, SSM, and hybrid models specifically for long-context inference on consumer and embedded GPUs. Our analysis reveals that SSMs are not only viable but superior for this domain, capable of processing sequences up to 220K tokens on a 24GB consumer GPU-approximately 4x longer than comparable Transformers. While Transformers may be up to 1.8x faster at short sequences, SSMs demonstrate a dramatic performance inversion, becoming up to 4x faster at very long contexts (~57K tokens). Our operator-level analysis reveals that custom, hardware-aware SSM kernels dominate the inference runtime, accounting for over 55% of latency on edge platforms, identifying them as a primary target for future hardware acceleration. We also provide detailed, device-specific characterization results to guide system co-design for the edge. To foster further research, we will open-source our characterization framework.
comment: 12 pages, 7 figures
☆ Describe Anything Model for Visual Question Answering on Text-rich Images ICCV 2025
Recent progress has been made in region-aware vision-language modeling, particularly with the emergence of the Describe Anything Model (DAM). DAM is capable of generating detailed descriptions of any specific image areas or objects without the need for additional localized image-text alignment supervision. We hypothesize that such region-level descriptive capability is beneficial for the task of Visual Question Answering (VQA), especially in challenging scenarios involving images with dense text. In such settings, the fine-grained extraction of textual information is crucial to producing correct answers. Motivated by this, we introduce DAM-QA, a framework with a tailored evaluation protocol, developed to investigate and harness the region-aware capabilities from DAM for the text-rich VQA problem that requires reasoning over text-based information within images. DAM-QA incorporates a mechanism that aggregates answers from multiple regional views of image content, enabling more effective identification of evidence that may be tied to text-related elements. Experiments on six VQA benchmarks show that our approach consistently outperforms the baseline DAM, with a notable 7+ point gain on DocVQA. DAM-QA also achieves the best overall performance among region-aware models with fewer parameters, significantly narrowing the gap with strong generalist VLMs. These results highlight the potential of DAM-like models for text-rich and broader VQA tasks when paired with efficient usage and integration strategies. Our code is publicly available at https://github.com/Linvyl/DAM-QA.git.
comment: 11 pages, 5 figures. Accepted to VisionDocs @ ICCV 2025
☆ EgoVLA: Learning Vision-Language-Action Models from Egocentric Human Videos
Real robot data collection for imitation learning has led to significant advancements in robotic manipulation. However, the requirement for robot hardware in the process fundamentally constrains the scale of the data. In this paper, we explore training Vision-Language-Action (VLA) models using egocentric human videos. The benefit of using human videos is not only for their scale but more importantly for the richness of scenes and tasks. With a VLA trained on human video that predicts human wrist and hand actions, we can perform Inverse Kinematics and retargeting to convert the human actions to robot actions. We fine-tune the model using a few robot manipulation demonstrations to obtain the robot policy, namely EgoVLA. We propose a simulation benchmark called Isaac Humanoid Manipulation Benchmark, where we design diverse bimanual manipulation tasks with demonstrations. We fine-tune and evaluate EgoVLA with Isaac Humanoid Manipulation Benchmark and show significant improvements over baselines and ablate the importance of human data. Videos can be found on our website: https://rchalyang.github.io/EgoVLA
comment: More videos can be found on our website: https://rchalyang.github.io/EgoVLA
☆ A Bayesian Incentive Mechanism for Poison-Resilient Federated Learning
Federated learning (FL) enables collaborative model training across decentralized clients while preserving data privacy. However, its open-participation nature exposes it to data-poisoning attacks, in which malicious actors submit corrupted model updates to degrade the global model. Existing defenses are often reactive, relying on statistical aggregation rules that can be computationally expensive and that typically assume an honest majority. This paper introduces a proactive, economic defense: a lightweight Bayesian incentive mechanism that makes malicious behavior economically irrational. Each training round is modeled as a Bayesian game of incomplete information in which the server, acting as the principal, uses a small, private validation dataset to verify update quality before issuing payments. The design satisfies Individual Rationality (IR) for benevolent clients, ensuring their participation is profitable, and Incentive Compatibility (IC), making poisoning an economically dominated strategy. Extensive experiments on non-IID partitions of MNIST and FashionMNIST demonstrate robustness: with 50% label-flipping adversaries on MNIST, the mechanism maintains 96.7% accuracy, only 0.3 percentage points lower than in a scenario with 30% label-flipping adversaries. This outcome is 51.7 percentage points better than standard FedAvg, which collapses under the same 50% attack. The mechanism is computationally light, budget-bounded, and readily integrates into existing FL frameworks, offering a practical route to economically robust and sustainable FL ecosystems.
☆ Targeted Deep Architectures: A TMLE-Based Framework for Robust Causal Inference in Neural Networks
Modern deep neural networks are powerful predictive tools yet often lack valid inference for causal parameters, such as treatment effects or entire survival curves. While frameworks like Double Machine Learning (DML) and Targeted Maximum Likelihood Estimation (TMLE) can debias machine-learning fits, existing neural implementations either rely on "targeted losses" that do not guarantee solving the efficient influence function equation or computationally expensive post-hoc "fluctuations" for multi-parameter settings. We propose Targeted Deep Architectures (TDA), a new framework that embeds TMLE directly into the network's parameter space with no restrictions on the backbone architecture. Specifically, TDA partitions model parameters - freezing all but a small "targeting" subset - and iteratively updates them along a targeting gradient, derived from projecting the influence functions onto the span of the gradients of the loss with respect to weights. This procedure yields plug-in estimates that remove first-order bias and produce asymptotically valid confidence intervals. Crucially, TDA easily extends to multi-dimensional causal estimands (e.g., entire survival curves) by merging separate targeting gradients into a single universal targeting update. Theoretically, TDA inherits classical TMLE properties, including double robustness and semiparametric efficiency. Empirically, on the benchmark IHDP dataset (average treatment effects) and simulated survival data with informative censoring, TDA reduces bias and improves coverage relative to both standard neural-network estimators and prior post-hoc approaches. In doing so, TDA establishes a direct, scalable pathway toward rigorous causal inference within modern deep architectures for complex multi-parameter targets.
☆ Can We Predict Alignment Before Models Finish Thinking? Towards Monitoring Misaligned Reasoning Models
Open-weights reasoning language models generate long chains-of-thought (CoTs) before producing a final response, which improves performance but introduces additional alignment risks, with harmful content often appearing in both the CoTs and the final outputs. In this work, we investigate if we can use CoTs to predict final response misalignment. We evaluate a range of monitoring approaches, including humans, highly-capable large language models, and text classifiers, using either CoT text or activations. First, we find that a simple linear probe trained on CoT activations can significantly outperform all text-based methods in predicting whether a final response will be safe or unsafe. CoT texts are often unfaithful and can mislead humans and classifiers, while model latents (i.e., CoT activations) offer a more reliable predictive signal. Second, the probe makes accurate predictions before reasoning completes, achieving strong performance even when applied to early CoT segments. These findings generalize across model sizes, families, and safety benchmarks, suggesting that lightweight probes could enable real-time safety monitoring and early intervention during generation.
☆ Unit-Based Histopathology Tissue Segmentation via Multi-Level Feature Representation
We propose UTS, a unit-based tissue segmentation framework for histopathology that classifies each fixed-size 32 * 32 tile, rather than each pixel, as the segmentation unit. This approach reduces annotation effort and improves computational efficiency without compromising accuracy. To implement this approach, we introduce a Multi-Level Vision Transformer (L-ViT), which benefits the multi-level feature representation to capture both fine-grained morphology and global tissue context. Trained to segment breast tissue into three categories (infiltrating tumor, non-neoplastic stroma, and fat), UTS supports clinically relevant tasks such as tumor-stroma quantification and surgical margin assessment. Evaluated on 386,371 tiles from 459 H&E-stained regions, it outperforms U-Net variants and transformer-based baselines. Code and Dataset will be available at GitHub.
comment: 12 pages, 6 figures
☆ Mixture of Raytraced Experts
We introduce a Mixture of Raytraced Experts, a stacked Mixture of Experts (MoE) architecture which can dynamically select sequences of experts, producing computational graphs of variable width and depth. Existing MoE architectures generally require a fixed amount of computation for a given sample. Our approach, in contrast, yields predictions with increasing accuracy as the computation cycles through the experts' sequence. We train our model by iteratively sampling from a set of candidate experts, unfolding the sequence akin to how Recurrent Neural Networks are trained. Our method does not require load-balancing mechanisms, and preliminary experiments show a reduction in training epochs of 10\% to 40\% with a comparable/higher accuracy. These results point to new research directions in the field of MoEs, allowing the design of potentially faster and more expressive models. The code is available at https://github.com/nutig/RayTracing
comment: Preliminary version (pre-submission)
☆ AutoVDC: Automated Vision Data Cleaning Using Vision-Language Models
Training of autonomous driving systems requires extensive datasets with precise annotations to attain robust performance. Human annotations suffer from imperfections, and multiple iterations are often needed to produce high-quality datasets. However, manually reviewing large datasets is laborious and expensive. In this paper, we introduce AutoVDC (Automated Vision Data Cleaning) framework and investigate the utilization of Vision-Language Models (VLMs) to automatically identify erroneous annotations in vision datasets, thereby enabling users to eliminate these errors and enhance data quality. We validate our approach using the KITTI and nuImages datasets, which contain object detection benchmarks for autonomous driving. To test the effectiveness of AutoVDC, we create dataset variants with intentionally injected erroneous annotations and observe the error detection rate of our approach. Additionally, we compare the detection rates using different VLMs and explore the impact of VLM fine-tuning on our pipeline. The results demonstrate our method's high performance in error detection and data cleaning experiments, indicating its potential to significantly improve the reliability and accuracy of large-scale production datasets in autonomous driving.
☆ NOCTA: Non-Greedy Objective Cost-Tradeoff Acquisition for Longitudinal Data
In many critical applications, resource constraints limit the amount of information that can be gathered to make predictions. For example, in healthcare, patient data often spans diverse features ranging from lab tests to imaging studies. Each feature may carry different information and must be acquired at a respective cost of time, money, or risk to the patient. Moreover, temporal prediction tasks, where both instance features and labels evolve over time, introduce additional complexity in deciding when or what information is important. In this work, we propose NOCTA, a Non-Greedy Objective Cost-Tradeoff Acquisition method that sequentially acquires the most informative features at inference time while accounting for both temporal dynamics and acquisition cost. We first introduce a cohesive estimation target for our NOCTA setting, and then develop two complementary estimators: 1) a non-parametric method based on nearest neighbors to guide the acquisition (NOCTA-NP), and 2) a parametric method that directly predicts the utility of potential acquisitions (NOCTA-P). Experiments on synthetic and real-world medical datasets demonstrate that both NOCTA variants outperform existing baselines.
☆ Neural Network-Guided Symbolic Regression for Interpretable Descriptor Discovery in Perovskite Catalysts
Understanding and predicting the activity of oxide perovskite catalysts for the oxygen evolution reaction (OER) requires descriptors that are both accurate and physically interpretable. While symbolic regression (SR) offers a path to discover such formulas, its performance degrades with high-dimensional inputs and small datasets. We present a two-phase framework that combines neural networks (NN), feature importance analysis, and symbolic regression (SR) to discover interpretable descriptors for OER activity in oxide perovskites. In Phase I, using a small dataset and seven structural features, we reproduce and improve the known {\mu}/t descriptor by engineering composite features and applying symbolic regression, achieving training and validation MAEs of 22.8 and 20.8 meV, respectively. In Phase II, we expand to 164 features, reduce dimensionality, and identify LUMO energy as a key electronic descriptor. A final formula using {\mu}/t, {\mu}/RA, and LUMO energy achieves improved accuracy (training and validation MAEs of 22.1 and 20.6 meV) with strong physical interpretability. Our results demonstrate that NN-guided symbolic regression enables accurate, interpretable, and physically meaningful descriptor discovery in data-scarce regimes, indicating interpretability need not sacrifice accuracy for materials informatics.
comment: 31 pages
☆ ROC-n-reroll: How verifier imperfection affects test-time scaling
Test-time scaling aims to improve language model performance by leveraging additional compute during inference. While many works have empirically studied techniques like Best-of-N (BoN) and rejection sampling that make use of a verifier to enable test-time scaling, there is little theoretical understanding of how verifier imperfection affects performance. In this work, we address this gap. Specifically, we prove how instance-level accuracy of these methods is precisely characterized by the geometry of the verifier's ROC curve. Interestingly, while scaling is determined by the local geometry of the ROC curve for rejection sampling, it depends on global properties of the ROC curve for BoN. As a consequence when the ROC curve is unknown, it is impossible to extrapolate the performance of rejection sampling based on the low-compute regime. Furthermore, while rejection sampling outperforms BoN for fixed compute, in the infinite-compute limit both methods converge to the same level of accuracy, determined by the slope of the ROC curve near the origin. Our theoretical results are confirmed by experiments on GSM8K using different versions of Llama and Qwen to generate and verify solutions.
comment: 35 pages, 9 Figures
☆ Trustworthy Tree-based Machine Learning by $MoS_2$ Flash-based Analog CAM with Inherent Soft Boundaries
The rapid advancement of artificial intelligence has raised concerns regarding its trustworthiness, especially in terms of interpretability and robustness. Tree-based models like Random Forest and XGBoost excel in interpretability and accuracy for tabular data, but scaling them remains computationally expensive due to poor data locality and high data dependence. Previous efforts to accelerate these models with analog content addressable memory (CAM) have struggled, due to the fact that the difficult-to-implement sharp decision boundaries are highly susceptible to device variations, which leads to poor hardware performance and vulnerability to adversarial attacks. This work presents a novel hardware-software co-design approach using $MoS_2$ Flash-based analog CAM with inherent soft boundaries, enabling efficient inference with soft tree-based models. Our soft tree model inference experiments on $MoS_2$ analog CAM arrays show this method achieves exceptional robustness against device variation and adversarial attacks while achieving state-of-the-art accuracy. Specifically, our fabricated analog CAM arrays achieve $96\%$ accuracy on Wisconsin Diagnostic Breast Cancer (WDBC) database, while maintaining decision explainability. Our experimentally calibrated model validated only a $0.6\%$ accuracy drop on the MNIST dataset under $10\%$ device threshold variation, compared to a $45.3\%$ drop for traditional decision trees. This work paves the way for specialized hardware that enhances AI's trustworthiness and efficiency.
☆ Improving Reinforcement Learning Sample-Efficiency using Local Approximation
In this study, we derive Probably Approximately Correct (PAC) bounds on the asymptotic sample-complexity for RL within the infinite-horizon Markov Decision Process (MDP) setting that are sharper than those in existing literature. The premise of our study is twofold: firstly, the further two states are from each other, transition-wise, the less relevant the value of the first state is when learning the $\epsilon$-optimal value of the second; secondly, the amount of 'effort', sample-complexity-wise, expended in learning the $\epsilon$-optimal value of a state is independent of the number of samples required to learn the $\epsilon$-optimal value of a second state that is a sufficient number of transitions away from the first. Inversely, states within each other's vicinity have values that are dependent on each other and will require a similar number of samples to learn. By approximating the original MDP using smaller MDPs constructed using subsets of the original's state-space, we are able to reduce the sample-complexity by a logarithmic factor to $O(SA \log A)$ timesteps, where $S$ and $A$ are the state and action space sizes. We are able to extend these results to an infinite-horizon, model-free setting by constructing a PAC-MDP algorithm with the aforementioned sample-complexity. We conclude with showing how significant the improvement is by comparing our algorithm against prior work in an experimental setting.
comment: Preprint
☆ Heat Kernel Goes Topological
Topological neural networks have emerged as powerful successors of graph neural networks. However, they typically involve higher-order message passing, which incurs significant computational expense. We circumvent this issue with a novel topological framework that introduces a Laplacian operator on combinatorial complexes (CCs), enabling efficient computation of heat kernels that serve as node descriptors. Our approach captures multiscale information and enables permutation-equivariant representations, allowing easy integration into modern transformer-based architectures. Theoretically, the proposed method is maximally expressive because it can distinguish arbitrary non-isomorphic CCs. Empirically, it significantly outperforms existing topological methods in terms of computational efficiency. Besides demonstrating competitive performance with the state-of-the-art descriptors on standard molecular datasets, it exhibits superior capability in distinguishing complex topological structures and avoiding blind spots on topological benchmarks. Overall, this work advances topological deep learning by providing expressive yet scalable representations, thereby opening up exciting avenues for molecular classification and property prediction tasks.
☆ Nonlinear Concept Erasure: a Density Matching Approach ECAI 2025
Ensuring that neural models used in real-world applications cannot infer sensitive information, such as demographic attributes like gender or race, from text representations is a critical challenge when fairness is a concern. We address this issue through concept erasure, a process that removes information related to a specific concept from distributed representations while preserving as much of the remaining semantic information as possible. Our approach involves learning an orthogonal projection in the embedding space, designed to make the class-conditional feature distributions of the discrete concept to erase indistinguishable after projection. By adjusting the rank of the projector, we control the extent of information removal, while its orthogonality ensures strict preservation of the local structure of the embeddings. Our method, termed $\overline{\mathrm{L}}$EOPARD, achieves state-of-the-art performance in nonlinear erasure of a discrete attribute on classic natural language processing benchmarks. Furthermore, we demonstrate that $\overline{\mathrm{L}}$EOPARD effectively mitigates bias in deep nonlinear classifiers, thereby promoting fairness.
comment: 17 pages, 10 figures, accepted for publication in ECAI 2025 (28th European Conference on Artificial Intelligence)
☆ Neural Polar Decoders for Deletion Channels
This paper introduces a neural polar decoder (NPD) for deletion channels with a constant deletion rate. Existing polar decoders for deletion channels exhibit high computational complexity of $O(N^4)$, where $N$ is the block length. This limits the application of polar codes for deletion channels to short-to-moderate block lengths. In this work, we demonstrate that employing NPDs for deletion channels can reduce the computational complexity. First, we extend the architecture of the NPD to support deletion channels. Specifically, the NPD architecture consists of four neural networks (NNs), each replicating fundamental successive cancellation (SC) decoder operations. To support deletion channels, we change the architecture of only one. The computational complexity of the NPD is $O(AN\log N)$, where the parameter $A$ represents a computational budget determined by the user and is independent of the channel. We evaluate the new extended NPD for deletion channels with deletion rates $\delta\in\{0.01, 0.1\}$ and we verify the NPD with the ground truth given by the trellis decoder by Tal et al. We further show that due to the reduced complexity of the NPD, we are able to incorporate list decoding and further improve performance. We believe that the extended NPD presented here could have applications in future technologies like DNA storage.
☆ Compositional Discrete Latent Code for High Fidelity, Productive Diffusion Models
We argue that diffusion models' success in modeling complex distributions is, for the most part, coming from their input conditioning. This paper investigates the representation used to condition diffusion models from the perspective that ideal representations should improve sample fidelity, be easy to generate, and be compositional to allow out-of-training samples generation. We introduce Discrete Latent Code (DLC), an image representation derived from Simplicial Embeddings trained with a self-supervised learning objective. DLCs are sequences of discrete tokens, as opposed to the standard continuous image embeddings. They are easy to generate and their compositionality enables sampling of novel images beyond the training distribution. Diffusion models trained with DLCs have improved generation fidelity, establishing a new state-of-the-art for unconditional image generation on ImageNet. Additionally, we show that composing DLCs allows the image generator to produce out-of-distribution samples that coherently combine the semantics of images in diverse ways. Finally, we showcase how DLCs can enable text-to-image generation by leveraging large-scale pretrained language models. We efficiently finetune a text diffusion language model to generate DLCs that produce novel samples outside of the image generator training distribution.
comment: In submission, 22 pages, 7 tables, 12 figures
☆ Thought Purity: Defense Paradigm For Chain-of-Thought Attack
While reinforcement learning-trained Large Reasoning Models (LRMs, e.g., Deepseek-R1) demonstrate advanced reasoning capabilities in the evolving Large Language Models (LLMs) domain, their susceptibility to security threats remains a critical vulnerability. This weakness is particularly evident in Chain-of-Thought (CoT) generation processes, where adversarial methods like backdoor prompt attacks can systematically subvert the model's core reasoning mechanisms. The emerging Chain-of-Thought Attack (CoTA) reveals this vulnerability through exploiting prompt controllability, simultaneously degrading both CoT safety and task performance with low-cost interventions. To address this compounded security-performance vulnerability, we propose Thought Purity (TP): a defense paradigm that systematically strengthens resistance to malicious content while preserving operational efficacy. Our solution achieves this through three synergistic components: (1) a safety-optimized data processing pipeline (2) reinforcement learning-enhanced rule constraints (3) adaptive monitoring metrics. Our approach establishes the first comprehensive defense mechanism against CoTA vulnerabilities in reinforcement learning-aligned reasoning systems, significantly advancing the security-functionality equilibrium for next-generation AI architectures.
☆ PROL : Rehearsal Free Continual Learning in Streaming Data via Prompt Online Learning ICCV 2025
The data privacy constraint in online continual learning (OCL), where the data can be seen only once, complicates the catastrophic forgetting problem in streaming data. A common approach applied by the current SOTAs in OCL is with the use of memory saving exemplars or features from previous classes to be replayed in the current task. On the other hand, the prompt-based approach performs excellently in continual learning but with the cost of a growing number of trainable parameters. The first approach may not be applicable in practice due to data openness policy, while the second approach has the issue of throughput associated with the streaming data. In this study, we propose a novel prompt-based method for online continual learning that includes 4 main components: (1) single light-weight prompt generator as a general knowledge, (2) trainable scaler-and-shifter as specific knowledge, (3) pre-trained model (PTM) generalization preserving, and (4) hard-soft updates mechanism. Our proposed method achieves significantly higher performance than the current SOTAs in CIFAR100, ImageNet-R, ImageNet-A, and CUB dataset. Our complexity analysis shows that our method requires a relatively smaller number of parameters and achieves moderate training time, inference time, and throughput. For further study, the source code of our method is available at https://github.com/anwarmaxsum/PROL.
comment: ICCV 2025
☆ RegCL: Continual Adaptation of Segment Anything Model via Model Merging
To address the performance limitations of the Segment Anything Model (SAM) in specific domains, existing works primarily adopt adapter-based one-step adaptation paradigms. However, some of these methods are specific developed for specific domains. If used on other domains may lead to performance degradation. This issue of catastrophic forgetting severely limits the model's scalability. To address this issue, this paper proposes RegCL, a novel non-replay continual learning (CL) framework designed for efficient multi-domain knowledge integration through model merging. Specifically, RegCL incorporates the model merging algorithm into the continual learning paradigm by merging the parameters of SAM's adaptation modules (e.g., LoRA modules) trained on different domains. The merging process is guided by weight optimization, which minimizes prediction discrepancies between the merged model and each of the domain-specific models. RegCL effectively consolidates multi-domain knowledge while maintaining parameter efficiency, i.e., the model size remains constant regardless of the number of tasks, and no historical data storage is required. Experimental results demonstrate that RegCL achieves favorable continual learning performance across multiple downstream datasets, validating its effectiveness in dynamic scenarios.
☆ Text-ADBench: Text Anomaly Detection Benchmark based on LLMs Embedding
Text anomaly detection is a critical task in natural language processing (NLP), with applications spanning fraud detection, misinformation identification, spam detection and content moderation, etc. Despite significant advances in large language models (LLMs) and anomaly detection algorithms, the absence of standardized and comprehensive benchmarks for evaluating the existing anomaly detection methods on text data limits rigorous comparison and development of innovative approaches. This work performs a comprehensive empirical study and introduces a benchmark for text anomaly detection, leveraging embeddings from diverse pre-trained language models across a wide array of text datasets. Our work systematically evaluates the effectiveness of embedding-based text anomaly detection by incorporating (1) early language models (GloVe, BERT); (2) multiple LLMs (LLaMa-2, LLama-3, Mistral, OpenAI (small, ada, large)); (3) multi-domain text datasets (news, social media, scientific publications); (4) comprehensive evaluation metrics (AUROC, AUPRC). Our experiments reveal a critical empirical insight: embedding quality significantly governs anomaly detection efficacy, and deep learning-based approaches demonstrate no performance advantage over conventional shallow algorithms (e.g., KNN, Isolation Forest) when leveraging LLM-derived embeddings.In addition, we observe strongly low-rank characteristics in cross-model performance matrices, which enables an efficient strategy for rapid model evaluation (or embedding evaluation) and selection in practical applications. Furthermore, by open-sourcing our benchmark toolkit that includes all embeddings from different models and code at https://github.com/jicongfan/Text-Anomaly-Detection-Benchmark, this work provides a foundation for future research in robust and scalable text anomaly detection systems.
☆ Site-Level Fine-Tuning with Progressive Layer Freezing: Towards Robust Prediction of Bronchopulmonary Dysplasia from Day-1 Chest Radiographs in Extremely Preterm Infants
Bronchopulmonary dysplasia (BPD) is a chronic lung disease affecting 35% of extremely low birth weight infants. Defined by oxygen dependence at 36 weeks postmenstrual age, it causes lifelong respiratory complications. However, preventive interventions carry severe risks, including neurodevelopmental impairment, ventilator-induced lung injury, and systemic complications. Therefore, early BPD prognosis and prediction of BPD outcome is crucial to avoid unnecessary toxicity in low risk infants. Admission radiographs of extremely preterm infants are routinely acquired within 24h of life and could serve as a non-invasive prognostic tool. In this work, we developed and investigated a deep learning approach using chest X-rays from 163 extremely low-birth-weight infants ($\leq$32 weeks gestation, 401-999g) obtained within 24 hours of birth. We fine-tuned a ResNet-50 pretrained specifically on adult chest radiographs, employing progressive layer freezing with discriminative learning rates to prevent overfitting and evaluated a CutMix augmentation and linear probing. For moderate/severe BPD outcome prediction, our best performing model with progressive freezing, linear probing and CutMix achieved an AUROC of 0.78 $\pm$ 0.10, balanced accuracy of 0.69 $\pm$ 0.10, and an F1-score of 0.67 $\pm$ 0.11. In-domain pre-training significantly outperformed ImageNet initialization (p = 0.031) which confirms domain-specific pretraining to be important for BPD outcome prediction. Routine IRDS grades showed limited prognostic value (AUROC 0.57 $\pm$ 0.11), confirming the need of learned markers. Our approach demonstrates that domain-specific pretraining enables accurate BPD prediction from routine day-1 radiographs. Through progressive freezing and linear probing, the method remains computationally feasible for site-level implementation and future federated learning deployments.
comment: S.G.-F., M.B., and A.E. contributed equally to this work and share first authorship. M.Z. and P.F. contributed equally to this work and share senior authorship
☆ A Framework for Nonstationary Gaussian Processes with Neural Network Parameters
Gaussian processes have become a popular tool for nonparametric regression because of their flexibility and uncertainty quantification. However, they often use stationary kernels, which limit the expressiveness of the model and may be unsuitable for many datasets. We propose a framework that uses nonstationary kernels whose parameters vary across the feature space, modeling these parameters as the output of a neural network that takes the features as input. The neural network and Gaussian process are trained jointly using the chain rule to calculate derivatives. Our method clearly describes the behavior of the nonstationary parameters and is compatible with approximation methods for scaling to large datasets. It is flexible and easily adapts to different nonstationary kernels without needing to redesign the optimization procedure. Our methods are implemented with the GPyTorch library and can be readily modified. We test a nonstationary variance and noise variant of our method on several machine learning datasets and find that it achieves better accuracy and log-score than both a stationary model and a hierarchical model approximated with variational inference. Similar results are observed for a model with only nonstationary variance. We also demonstrate our approach's ability to recover the nonstationary parameters of a spatial dataset.
☆ Robust Causal Discovery in Real-World Time Series with Power-Laws
Exploring causal relationships in stochastic time series is a challenging yet crucial task with a vast range of applications, including finance, economics, neuroscience, and climate science. Many algorithms for Causal Discovery (CD) have been proposed, but they often exhibit a high sensitivity to noise, resulting in misleading causal inferences when applied to real data. In this paper, we observe that the frequency spectra of typical real-world time series follow a power-law distribution, notably due to an inherent self-organizing behavior. Leveraging this insight, we build a robust CD method based on the extraction of power -law spectral features that amplify genuine causal signals. Our method consistently outperforms state-of-the-art alternatives on both synthetic benchmarks and real-world datasets with known causal structures, demonstrating its robustness and practical relevance.
☆ Surrogate Quantum Circuit Design for the Lattice Boltzmann Collision Operator
Direct numerical simulation of turbulent flows at high Reynolds numbers remains a major challenge for traditional computational fluid dynamics (CFD) tools running on classical computer hardware. This has motivated growing interest in quantum algorithms for CFD to enable flow simulations on quantum computers. The reason being that these computers are expected to deliver potential speed-ups for certain problems. One promising quantum CFD approach is a fully quantum implementation of the lattice Boltzmann method called QLBM. Although efficient quantum routines are now available for the streaming step, implementing the nonlinear, irreversible collision step with a low depth circuit that avoids additional ancilla qubits, probabilistic post-selection and repeated executions remains a significant challenge. In this study, we address this challenge by introducing a framework for learning a surrogate quantum circuit (SQC) that approximates the full Bhatnagar Gross Krook (BGK) collision operator for the D2Q9 lattice. The four qubit circuit is trained to respect the physical properties of the BGK collision operator, including mass and momentum conservation, D8 equivariance and scale equivariance. When compiled to the gate set used by IBM Heron processor under the assumption of full qubit connectivity, the 15 block SQC requires only 2,430 native gates and uses neither ancilla qubits nor post-selection or repeated executions. Moreover, its depth is independent of the grid resolution, as collision is a local operation that can exploit quantum parallelism to its full extent. We validate the SQC on two benchmark flows, the Taylor Green vortex decay and the lid driven cavity, demonstrating that it accurately captures vortex dissipation and flow recirculation.
comment: 31 pages, 14 figures
☆ Comparative Analysis of CNN Performance in Keras, PyTorch and JAX on PathMNIST
Deep learning has significantly advanced the field of medical image classification, particularly with the adoption of Convolutional Neural Networks (CNNs). Various deep learning frameworks such as Keras, PyTorch and JAX offer unique advantages in model development and deployment. However, their comparative performance in medical imaging tasks remains underexplored. This study presents a comprehensive analysis of CNN implementations across these frameworks, using the PathMNIST dataset as a benchmark. We evaluate training efficiency, classification accuracy and inference speed to assess their suitability for real-world applications. Our findings highlight the trade-offs between computational speed and model accuracy, offering valuable insights for researchers and practitioners in medical image analysis.
☆ Universal Fourier Neural Operators for Micromechanics
\noindent Solving cell problems in homogenization is hard, and available deep-learning frameworks fail to match the speed and generality of traditional computational frameworks. More to the point, it is generally unclear what to expect of machine-learning approaches, let alone single out which approaches are promising. In the work at hand, we advocate Fourier Neural Operators (FNOs) for micromechanics, empowering them by insights from computational micromechanics methods based on the fast Fourier transform (FFT). We construct an FNO surrogate mimicking the basic scheme foundational for FFT-based methods and show that the resulting operator predicts solutions to cell problems with \emph{arbitrary} stiffness distribution only subject to a material-contrast constraint up to a desired accuracy. In particular, there are no restrictions on the material symmetry like isotropy, on the number of phases and on the geometry of the interfaces between materials. Also, the provided fidelity is sharp and uniform, providing explicit guarantees leveraging our physical empowerment of FNOs. To show the desired universal approximation property, we construct an FNO explicitly that requires no training to begin with. Still, the obtained neural operator complies with the same memory requirements as the basic scheme and comes with runtimes proportional to classical FFT solvers. In particular, large-scale problems with more than 100 million voxels are readily handled. The goal of this work is to underline the potential of FNOs for solving micromechanical problems, linking FFT-based methods to FNOs. This connection is expected to provide a fruitful exchange between both worlds.
comment: 48 pages, 13 figures
☆ Optimizers Qualitatively Alter Solutions And We Should Leverage This
Due to the nonlinear nature of Deep Neural Networks (DNNs), one can not guarantee convergence to a unique global minimum of the loss when using optimizers relying only on local information, such as SGD. Indeed, this was a primary source of skepticism regarding the feasibility of DNNs in the early days of the field. The past decades of progress in deep learning have revealed this skepticism to be misplaced, and a large body of empirical evidence shows that sufficiently large DNNs following standard training protocols exhibit well-behaved optimization dynamics that converge to performant solutions. This success has biased the community to use convex optimization as a mental model for learning, leading to a focus on training efficiency, either in terms of required iteration, FLOPs or wall-clock time, when improving optimizers. We argue that, while this perspective has proven extremely fruitful, another perspective specific to DNNs has received considerably less attention: the optimizer not only influences the rate of convergence, but also the qualitative properties of the learned solutions. Restated, the optimizer can and will encode inductive biases and change the effective expressivity of a given class of models. Furthermore, we believe the optimizer can be an effective way of encoding desiderata in the learning process. We contend that the community should aim at understanding the biases of already existing methods, as well as aim to build new optimizers with the explicit intent of inducing certain properties of the solution, rather than solely judging them based on their convergence rates. We hope our arguments will inspire research to improve our understanding of how the learning process can impact the type of solution we converge to, and lead to a greater recognition of optimizers design as a critical lever that complements the roles of architecture and data in shaping model outcomes.
☆ Physics-Informed Linear Model (PILM): Analytical Representations and Application to Crustal Strain Rate Estimation
Many physical systems are described by partial differential equations (PDEs), and solving these equations and estimating their coefficients or boundary conditions (BCs) from observational data play a crucial role in understanding the associated phenomena. Recently, a machine learning approach known as physics-informed neural network, which solves PDEs using neural networks by minimizing the sum of residuals from the PDEs, BCs, and data, has gained significant attention in the scientific community. In this study, we investigate a physics-informed linear model (PILM) that uses linear combinations of basis functions to represent solutions, thereby enabling an analytical representation of optimal solutions. The PILM was formulated and verified for illustrative forward and inverse problems including cases with uncertain BCs. Furthermore, the PILM was applied to estimate crustal strain rates using geodetic data. Specifically, physical regularization that enforces elastic equilibrium on the velocity fields was compared with mathematical regularization that imposes smoothness constraints. From a Bayesian perspective, mathematical regularization exhibited superior performance. The PILM provides an analytically solvable framework applicable to linear forward and inverse problems, underdetermined systems, and physical regularization.
☆ Sparse Autoencoders for Sequential Recommendation Models: Interpretation and Flexible Control
Many current state-of-the-art models for sequential recommendations are based on transformer architectures. Interpretation and explanation of such black box models is an important research question, as a better understanding of their internals can help understand, influence, and control their behavior, which is very important in a variety of real-world applications. Recently sparse autoencoders (SAE) have been shown to be a promising unsupervised approach for extracting interpretable features from language models. These autoencoders learn to reconstruct hidden states of the transformer's internal layers from sparse linear combinations of directions in their activation space. This paper is focused on the application of SAE to the sequential recommendation domain. We show that this approach can be successfully applied to the transformer trained on a sequential recommendation task: learned directions turn out to be more interpretable and monosemantic than the original hidden state dimensions. Moreover, we demonstrate that the features learned by SAE can be used to effectively and flexibly control the model's behavior, providing end-users with a straightforward method to adjust their recommendations to different custom scenarios and contexts.
☆ Selective Quantization Tuning for ONNX Models
Quantization is a process that reduces the precision of deep neural network models to lower model size and computational demands, often at the cost of accuracy. However, fully quantized models may exhibit sub-optimal performance below acceptable levels and face deployment challenges on low-end hardware accelerators due to practical constraints. To address these issues, quantization can be selectively applied to only a subset of layers, but selecting which layers to exclude is non-trivial. To this direction, we propose TuneQn, a suite enabling selective quantization, deployment and execution of ONNX models across various CPU and GPU devices, combined with profiling and multi-objective optimization. TuneQn generates selectively quantized ONNX models, deploys them on different hardware, measures performance on metrics like accuracy and size, performs Pareto Front minimization to identify the best model candidate and visualizes the results. To demonstrate the effectiveness of TuneQn, we evaluated TuneQn on four ONNX models with two quantization settings across CPU and GPU devices. As a result, we demonstrated that our utility effectively performs selective quantization and tuning, selecting ONNX model candidates with up to a $54.14$% reduction in accuracy loss compared to the fully quantized model, and up to a $72.9$% model size reduction compared to the original model.
comment: 5 pages, 3 figures, 2 tables
☆ Explainable Evidential Clustering
Unsupervised classification is a fundamental machine learning problem. Real-world data often contain imperfections, characterized by uncertainty and imprecision, which are not well handled by traditional methods. Evidential clustering, based on Dempster-Shafer theory, addresses these challenges. This paper explores the underexplored problem of explaining evidential clustering results, which is crucial for high-stakes domains such as healthcare. Our analysis shows that, in the general case, representativity is a necessary and sufficient condition for decision trees to serve as abductive explainers. Building on the concept of representativity, we generalize this idea to accommodate partial labeling through utility functions. These functions enable the representation of "tolerable" mistakes, leading to the definition of evidential mistakeness as explanation cost and the construction of explainers tailored to evidential classifiers. Finally, we propose the Iterative Evidential Mistake Minimization (IEMM) algorithm, which provides interpretable and cautious decision tree explanations for evidential clustering functions. We validate the proposed algorithm on synthetic and real-world data. Taking into account the decision-maker's preferences, we were able to provide an explanation that was satisfactory up to 93% of the time.
☆ BenchRL-QAS: Benchmarking reinforcement learning algorithms for quantum architecture search
We introduce BenchRL-QAS, a unified benchmarking framework for systematically evaluating reinforcement learning (RL) algorithms in quantum architecture search (QAS) across diverse variational quantum algorithm tasks and system sizes ranging from 2- to 8-qubit. Our study benchmarks nine RL agents including both value-based and policy-gradient methods on representative quantum problems such as variational quantum eigensolver, variational quantum state diagonalization, quantum classification, and state preparation, spanning both noiseless and realistic noisy regimes. We propose a weighted ranking metric that balances accuracy, circuit depth, gate count, and computational efficiency, enabling fair and comprehensive comparison. Our results first reveal that RL-based quantum classifier outperforms baseline variational classifiers. Then we conclude that no single RL algorithm is universally optimal when considering a set of QAS tasks; algorithmic performance is highly context-dependent, varying with task structure, qubit count, and noise. This empirical finding provides strong evidence for the "no free lunch" principle in RL-based quantum circuit design and highlights the necessity of tailored algorithm selection and systematic benchmarking for advancing quantum circuit synthesis. This work represents the most comprehensive RL-QAS benchmarking effort to date, and BenchRL-QAS along with all experimental data are made publicly available to support reproducibility and future research https://github.com/azhar-ikhtiarudin/bench-rlqas.
comment: Comprehensive RL agent benchmark for QAS. Contributions are welcomed here: https://github.com/azhar-ikhtiarudin/bench-rlqas
☆ RUMAA: Repeat-Aware Unified Music Audio Analysis for Score-Performance Alignment, Transcription, and Mistake Detection SP
This study introduces RUMAA, a transformer-based framework for music performance analysis that unifies score-to-performance alignment, score-informed transcription, and mistake detection in a near end-to-end manner. Unlike prior methods addressing these tasks separately, RUMAA integrates them using pre-trained score and audio encoders and a novel tri-stream decoder capturing task interdependencies through proxy tasks. It aligns human-readable MusicXML scores with repeat symbols to full-length performance audio, overcoming traditional MIDI-based methods that rely on manually unfolded score-MIDI data with pre-specified repeat structures. RUMAA matches state-of-the-art alignment methods on non-repeated scores and outperforms them on scores with repeats in a public piano music dataset, while also delivering promising transcription and mistake detection results.
comment: Accepted to WASPAA 2025
☆ RadioDiff-3D: A 3D$\times$3D Radio Map Dataset and Generative Diffusion Based Benchmark for 6G Environment-Aware Communication
Radio maps (RMs) serve as a critical foundation for enabling environment-aware wireless communication, as they provide the spatial distribution of wireless channel characteristics. Despite recent progress in RM construction using data-driven approaches, most existing methods focus solely on pathloss prediction in a fixed 2D plane, neglecting key parameters such as direction of arrival (DoA), time of arrival (ToA), and vertical spatial variations. Such a limitation is primarily due to the reliance on static learning paradigms, which hinder generalization beyond the training data distribution. To address these challenges, we propose UrbanRadio3D, a large-scale, high-resolution 3D RM dataset constructed via ray tracing in realistic urban environments. UrbanRadio3D is over 37$\times$3 larger than previous datasets across a 3D space with 3 metrics as pathloss, DoA, and ToA, forming a novel 3D$\times$33D dataset with 7$\times$3 more height layers than prior state-of-the-art (SOTA) dataset. To benchmark 3D RM construction, a UNet with 3D convolutional operators is proposed. Moreover, we further introduce RadioDiff-3D, a diffusion-model-based generative framework utilizing the 3D convolutional architecture. RadioDiff-3D supports both radiation-aware scenarios with known transmitter locations and radiation-unaware settings based on sparse spatial observations. Extensive evaluations on UrbanRadio3D validate that RadioDiff-3D achieves superior performance in constructing rich, high-dimensional radio maps under diverse environmental dynamics. This work provides a foundational dataset and benchmark for future research in 3D environment-aware communication. The dataset is available at https://github.com/UNIC-Lab/UrbanRadio3D.
☆ Multi-Component VAE with Gaussian Markov Random Field
Multi-component datasets with intricate dependencies, like industrial assemblies or multi-modal imaging, challenge current generative modeling techniques. Existing Multi-component Variational AutoEncoders typically rely on simplified aggregation strategies, neglecting critical nuances and consequently compromising structural coherence across generated components. To explicitly address this gap, we introduce the Gaussian Markov Random Field Multi-Component Variational AutoEncoder , a novel generative framework embedding Gaussian Markov Random Fields into both prior and posterior distributions. This design choice explicitly models cross-component relationships, enabling richer representation and faithful reproduction of complex interactions. Empirically, our GMRF MCVAE achieves state-of-the-art performance on a synthetic Copula dataset specifically constructed to evaluate intricate component relationships, demonstrates competitive results on the PolyMNIST benchmark, and significantly enhances structural coherence on the real-world BIKED dataset. Our results indicate that the GMRF MCVAE is especially suited for practical applications demanding robust and realistic modeling of multi-component coherence
☆ PRISM: Distributed Inference for Foundation Models at Edge
Foundation models (FMs) have achieved remarkable success across a wide range of applications, from image classification to natural langurage processing, but pose significant challenges for deployment at edge. This has sparked growing interest in developing practical and efficient strategies for bringing foundation models to edge environments. In this work, we propose PRISM, a communication-efficient and compute-aware strategy for distributed Transformer inference on edge devices. Our method leverages a Segment Means representation to approximate intermediate output features, drastically reducing inter-device communication. Additionally, we restructure the self-attention mechanism to eliminate redundant computations caused by per-device Key/Value calculation in position-wise partitioning and design a partition-aware causal masking scheme tailored for autoregressive models. We evaluate PRISM on ViT, BERT, and GPT-2 across diverse datasets, namely CIFAR-10, CIFAR-100, ImageNet-1k, GLUE, and CBT. Our results demonstrate substantial reductions in communication overhead (up to 99.2% for BERT at compression rate CR = 128) and per-device computation (51.24% for BERT at the same setting), with only minor accuracy degradation. This method offers a scalable and practical solution for deploying foundation models in distributed resource-constrained environments.
☆ FourCastNet 3: A geometric approach to probabilistic machine-learning weather forecasting at scale
FourCastNet 3 advances global weather modeling by implementing a scalable, geometric machine learning (ML) approach to probabilistic ensemble forecasting. The approach is designed to respect spherical geometry and to accurately model the spatially correlated probabilistic nature of the problem, resulting in stable spectra and realistic dynamics across multiple scales. FourCastNet 3 delivers forecasting accuracy that surpasses leading conventional ensemble models and rivals the best diffusion-based methods, while producing forecasts 8 to 60 times faster than these approaches. In contrast to other ML approaches, FourCastNet 3 demonstrates excellent probabilistic calibration and retains realistic spectra, even at extended lead times of up to 60 days. All of these advances are realized using a purely convolutional neural network architecture tailored for spherical geometry. Scalable and efficient large-scale training on 1024 GPUs and more is enabled by a novel training paradigm for combined model- and data-parallelism, inspired by domain decomposition methods in classical numerical models. Additionally, FourCastNet 3 enables rapid inference on a single GPU, producing a 90-day global forecast at 0.25{\deg}, 6-hourly resolution in under 20 seconds. Its computational efficiency, medium-range probabilistic skill, spectral fidelity, and rollout stability at subseasonal timescales make it a strong candidate for improving meteorological forecasting and early warning systems through large ensemble predictions.
☆ RiemannLoRA: A Unified Riemannian Framework for Ambiguity-Free LoRA Optimization
Low-Rank Adaptation (LoRA) has become a widely adopted standard for parameter-efficient fine-tuning of large language models (LLMs), significantly reducing memory and computational demands. However, challenges remain, including finding optimal initialization strategies or mitigating overparametrization in low-rank matrix factorization. In this work, we propose a novel approach that addresses both of the challenges simultaneously within a unified framework. Our method treats a set of fixed-rank LoRA matrices as a smooth manifold. Considering adapters as elements on this manifold removes overparametrization, while determining the direction of the fastest loss decrease along the manifold provides initialization. Special care is taken to obtain numerically stable and computationally efficient implementation of our method, using best practices from numerical linear algebra and Riemannian optimization. Experimental results on LLM and diffusion model architectures demonstrate that RiemannLoRA consistently improves both convergence speed and final performance over standard LoRA and its state-of-the-art modifications.
☆ Neural Human Pose Prior
We introduce a principled, data-driven approach for modeling a neural prior over human body poses using normalizing flows. Unlike heuristic or low-expressivity alternatives, our method leverages RealNVP to learn a flexible density over poses represented in the 6D rotation format. We address the challenge of modeling distributions on the manifold of valid 6D rotations by inverting the Gram-Schmidt process during training, enabling stable learning while preserving downstream compatibility with rotation-based frameworks. Our architecture and training pipeline are framework-agnostic and easily reproducible. We demonstrate the effectiveness of the learned prior through both qualitative and quantitative evaluations, and we analyze its impact via ablation studies. This work provides a sound probabilistic foundation for integrating pose priors into human motion capture and reconstruction pipelines.
comment: Work in progress
☆ HyDRA: A Hybrid Dual-Mode Network for Closed- and Open-Set RFFI with Optimized VMD
Device recognition is vital for security in wireless communication systems, particularly for applications like access control. Radio Frequency Fingerprint Identification (RFFI) offers a non-cryptographic solution by exploiting hardware-induced signal distortions. This paper proposes HyDRA, a Hybrid Dual-mode RF Architecture that integrates an optimized Variational Mode Decomposition (VMD) with a novel architecture based on the fusion of Convolutional Neural Networks (CNNs), Transformers, and Mamba components, designed to support both closed-set and open-set classification tasks. The optimized VMD enhances preprocessing efficiency and classification accuracy by fixing center frequencies and using closed-form solutions. HyDRA employs the Transformer Dynamic Sequence Encoder (TDSE) for global dependency modeling and the Mamba Linear Flow Encoder (MLFE) for linear-complexity processing, adapting to varying conditions. Evaluation on public datasets demonstrates state-of-the-art (SOTA) accuracy in closed-set scenarios and robust performance in our proposed open-set classification method, effectively identifying unauthorized devices. Deployed on NVIDIA Jetson Xavier NX, HyDRA achieves millisecond-level inference speed with low power consumption, providing a practical solution for real-time wireless authentication in real-world environments.
☆ Self-Adaptive and Robust Federated Spectrum Sensing without Benign Majority for Cellular Networks
Advancements in wireless and mobile technologies, including 5G advanced and the envisioned 6G, are driving exponential growth in wireless devices. However, this rapid expansion exacerbates spectrum scarcity, posing a critical challenge. Dynamic spectrum allocation (DSA)--which relies on sensing and dynamically sharing spectrum--has emerged as an essential solution to address this issue. While machine learning (ML) models hold significant potential for improving spectrum sensing, their adoption in centralized ML-based DSA systems is limited by privacy concerns, bandwidth constraints, and regulatory challenges. To overcome these limitations, distributed ML-based approaches such as Federated Learning (FL) offer promising alternatives. This work addresses two key challenges in FL-based spectrum sensing (FLSS). First, the scarcity of labeled data for training FL models in practical spectrum sensing scenarios is tackled with a semi-supervised FL approach, combined with energy detection, enabling model training on unlabeled datasets. Second, we examine the security vulnerabilities of FLSS, focusing on the impact of data poisoning attacks. Our analysis highlights the shortcomings of existing majority-based defenses in countering such attacks. To address these vulnerabilities, we propose a novel defense mechanism inspired by vaccination, which effectively mitigates data poisoning attacks without relying on majority-based assumptions. Extensive experiments on both synthetic and real-world datasets validate our solutions, demonstrating that FLSS can achieve near-perfect accuracy on unlabeled datasets and maintain Byzantine robustness against both targeted and untargeted data poisoning attacks, even when a significant proportion of participants are malicious.
☆ Iterative Augmentation with Summarization Refinement (IASR) Evaluation for Unstructured Survey data Modeling and Analysis
Text data augmentation is a widely used strategy for mitigating data sparsity in natural language processing (NLP), particularly in low-resource settings where limited samples hinder effective semantic modeling. While augmentation can improve input diversity and downstream interpretability, existing techniques often lack mechanisms to ensure semantic preservation during large-scale or iterative generation, leading to redundancy and instability. This work introduces a principled evaluation framework for large language model (LLM) based text augmentation, comprising two components: (1) Scalability Analysis, which measures semantic consistency as augmentation volume increases, and (2) Iterative Augmentation with Summarization Refinement (IASR), which evaluates semantic drift across recursive paraphrasing cycles. Empirical evaluations across state-of-the-art LLMs show that GPT-3.5 Turbo achieved the best balance of semantic fidelity, diversity, and generation efficiency. Applied to a real-world topic modeling task using BERTopic with GPT-enhanced few-shot labeling, the proposed approach results in a 400% increase in topic granularity and complete elimination of topic overlaps. These findings validated the utility of the proposed frameworks for structured evaluation of LLM-based augmentation in practical NLP pipelines.
☆ Multimodal Coordinated Online Behavior: Trade-offs and Strategies
Coordinated online behavior, which spans from beneficial collective actions to harmful manipulation such as disinformation campaigns, has become a key focus in digital ecosystem analysis. Traditional methods often rely on monomodal approaches, focusing on single types of interactions like co-retweets or co-hashtags, or consider multiple modalities independently of each other. However, these approaches may overlook the complex dynamics inherent in multimodal coordination. This study compares different ways of operationalizing the detection of multimodal coordinated behavior. It examines the trade-off between weakly and strongly integrated multimodal models, highlighting the balance between capturing broader coordination patterns and identifying tightly coordinated behavior. By comparing monomodal and multimodal approaches, we assess the unique contributions of different data modalities and explore how varying implementations of multimodality impact detection outcomes. Our findings reveal that not all the modalities provide distinct insights, but that with a multimodal approach we can get a more comprehensive understanding of coordination dynamics. This work enhances the ability to detect and analyze coordinated online behavior, offering new perspectives for safeguarding the integrity of digital platforms.
☆ A Privacy-Preserving Framework for Advertising Personalization Incorporating Federated Learning and Differential Privacy
To mitigate privacy leakage and performance issues in personalized advertising, this paper proposes a framework that integrates federated learning and differential privacy. The system combines distributed feature extraction, dynamic privacy budget allocation, and robust model aggregation to balance model accuracy, communication overhead, and privacy protection. Multi-party secure computing and anomaly detection mechanisms further enhance system resilience against malicious attacks. Experimental results demonstrate that the framework achieves dual optimization of recommendation accuracy and system efficiency while ensuring privacy, providing both a practical solution and a theoretical foundation for applying privacy protection technologies in advertisement recommendation.
☆ Measuring Informativeness Gap of (Mis)Calibrated Predictors
In many applications, decision-makers must choose between multiple predictive models that may all be miscalibrated. Which model (i.e., predictor) is more "useful" in downstream decision tasks? To answer this, our first contribution introduces the notion of the informativeness gap between any two predictors, defined as the maximum normalized payoff advantage one predictor offers over the other across all decision-making tasks. Our framework strictly generalizes several existing notions: it subsumes U-Calibration [KLST-23] and Calibration Decision Loss [HW-24], which compare a miscalibrated predictor to its calibrated counterpart, and it recovers Blackwell informativeness [Bla-51, Bla-53] as a special case when both predictors are perfectly calibrated. Our second contribution is a dual characterization of the informativeness gap, which gives rise to a natural informativeness measure that can be viewed as a relaxed variant of the earth mover's distance (EMD) between two prediction distributions. We show that this measure satisfies natural desiderata: it is complete and sound, and it can be estimated sample-efficiently in the prediction-only access setting. Along the way, we also obtain novel combinatorial structural results when applying this measure to perfectly calibrated predictors.
☆ Improved Analysis for Sign-based Methods with Momentum Updates
In this paper, we present enhanced analysis for sign-based optimization algorithms with momentum updates. Traditional sign-based methods, under the separable smoothness assumption, guarantee a convergence rate of $\mathcal{O}(T^{-1/4})$, but they either require large batch sizes or assume unimodal symmetric stochastic noise. To address these limitations, we demonstrate that signSGD with momentum can achieve the same convergence rate using constant batch sizes without additional assumptions. Our analysis, under the standard $l_2$-smoothness condition, improves upon the result of the prior momentum-based signSGD method by a factor of $\mathcal{O}(d^{1/2})$, where $d$ is the problem dimension. Furthermore, we explore sign-based methods with majority vote in distributed settings and show that the proposed momentum-based method yields convergence rates of $\mathcal{O}\left( d^{1/2}T^{-1/2} + dn^{-1/2} \right)$ and $\mathcal{O}\left( \max \{ d^{1/4}T^{-1/4}, d^{1/10}T^{-1/5} \} \right)$, which outperform the previous results of $\mathcal{O}\left( dT^{-1/4} + dn^{-1/2} \right)$ and $\mathcal{O}\left( d^{3/8}T^{-1/8} \right)$, respectively. Numerical experiments further validate the effectiveness of the proposed methods.
☆ Emergence of Quantised Representations Isolated to Anisotropic Functions
This paper describes a novel methodology for determining representational alignment, developed upon the existing Spotlight Resonance method. Using this, it is found that algebraic symmetries of network primitives are a strong predictor for task-agnostic structure in representations. Particularly, this new tool is used to gain insight into how discrete representations can form and arrange in autoencoder models, through an ablation study where only the activation function is altered. Representations are found to tend to discretise when the activation functions are defined through a discrete algebraic permutation-equivariant symmetry. In contrast, they remain continuous under a continuous algebraic orthogonal-equivariant definition. These findings corroborate the hypothesis that functional form choices can carry unintended inductive biases which produce task-independent artefactual structures in representations, particularly that contemporary forms induce discretisation of otherwise continuous structure -- a quantisation effect. Moreover, this supports a general causal model for one mode in which discrete representations may form, and could constitute a prerequisite for downstream interpretability phenomena, including grandmother neurons, discrete coding schemes, general linear features and possibly Superposition. Hence, this tool and proposed mechanism for the influence of functional form on representations may provide several insights into emergent interpretability research. Finally, preliminary results indicate that quantisation of representations appears to correlate with a measurable increase in reconstruction error, reinforcing previous conjectures that this collapse can be detrimental.
comment: 36 pages, 31 figures
☆ StylOch at PAN: Gradient-Boosted Trees with Frequency-Based Stylometric Features
This submission to the binary AI detection task is based on a modular stylometric pipeline, where: public spaCy models are used for text preprocessing (including tokenisation, named entity recognition, dependency parsing, part-of-speech tagging, and morphology annotation) and extracting several thousand features (frequencies of n-grams of the above linguistic annotations); light-gradient boosting machines are used as the classifier. We collect a large corpus of more than 500 000 machine-generated texts for the classifier's training. We explore several parameter options to increase the classifier's capacity and take advantage of that training set. Our approach follows the non-neural, computationally inexpensive but explainable approach found effective previously.
☆ FloGAN: Scenario-Based Urban Mobility Flow Generation via Conditional GANs and Dynamic Region Decoupling
The mobility patterns of people in cities evolve alongside changes in land use and population. This makes it crucial for urban planners to simulate and analyze human mobility patterns for purposes such as transportation optimization and sustainable urban development. Existing generative models borrowed from machine learning rely heavily on historical trajectories and often overlook evolving factors like changes in population density and land use. Mechanistic approaches incorporate population density and facility distribution but assume static scenarios, limiting their utility for future projections where historical data for calibration is unavailable. This study introduces a novel, data-driven approach for generating origin-destination mobility flows tailored to simulated urban scenarios. Our method leverages adaptive factors such as dynamic region sizes and land use archetypes, and it utilizes conditional generative adversarial networks (cGANs) to blend historical data with these adaptive parameters. The approach facilitates rapid mobility flow generation with adjustable spatial granularity based on regions of interest, without requiring extensive calibration data or complex behavior modeling. The promising performance of our approach is demonstrated by its application to mobile phone data from Singapore, and by its comparison with existing methods.
comment: International Conference on Intelligent Digitization of Systems and Services, Valencia, Spain, 2025 (IDSS 2025)
☆ Information-Theoretic Generalization Bounds of Replay-based Continual Learning
Continual learning (CL) has emerged as a dominant paradigm for acquiring knowledge from sequential tasks while avoiding catastrophic forgetting. Although many CL methods have been proposed to show impressive empirical performance, the theoretical understanding of their generalization behavior remains limited, particularly for replay-based approaches. In this paper, we establish a unified theoretical framework for replay-based CL, deriving a series of information-theoretic bounds that explicitly characterize how the memory buffer interacts with the current task to affect generalization. Specifically, our hypothesis-based bounds reveal that utilizing the limited exemplars of previous tasks alongside the current task data, rather than exhaustive replay, facilitates improved generalization while effectively mitigating catastrophic forgetting. Furthermore, our prediction-based bounds yield tighter and computationally tractable upper bounds of the generalization gap through the use of low-dimensional variables. Our analysis is general and broadly applicable to a wide range of learning algorithms, exemplified by stochastic gradient Langevin dynamics (SGLD) as a representative method. Comprehensive experimental evaluations demonstrate the effectiveness of our derived bounds in capturing the generalization dynamics in replay-based CL settings.
☆ Granular feedback merits sophisticated aggregation
Human feedback is increasingly used across diverse applications like training AI models, developing recommender systems, and measuring public opinion -- with granular feedback often being preferred over binary feedback for its greater informativeness. While it is easy to accurately estimate a population's distribution of feedback given feedback from a large number of individuals, cost constraints typically necessitate using smaller groups. A simple method to approximate the population distribution is regularized averaging: compute the empirical distribution and regularize it toward a prior. Can we do better? As we will discuss, the answer to this question depends on feedback granularity. Suppose one wants to predict a population's distribution of feedback using feedback from a limited number of individuals. We show that, as feedback granularity increases, one can substantially improve upon predictions of regularized averaging by combining individuals' feedback in ways more sophisticated than regularized averaging. Our empirical analysis using questions on social attitudes confirms this pattern. In particular, with binary feedback, sophistication barely reduces the number of individuals required to attain a fixed level of performance. By contrast, with five-point feedback, sophisticated methods match the performance of regularized averaging with about half as many individuals.
comment: 31 pages, 8 figures
☆ MVAR: MultiVariate AutoRegressive Air Pollutants Forecasting Model
Air pollutants pose a significant threat to the environment and human health, thus forecasting accurate pollutant concentrations is essential for pollution warnings and policy-making. Existing studies predominantly focus on single-pollutant forecasting, neglecting the interactions among different pollutants and their diverse spatial responses. To address the practical needs of forecasting multivariate air pollutants, we propose MultiVariate AutoRegressive air pollutants forecasting model (MVAR), which reduces the dependency on long-time-window inputs and boosts the data utilization efficiency. We also design the Multivariate Autoregressive Training Paradigm, enabling MVAR to achieve 120-hour long-term sequential forecasting. Additionally, MVAR develops Meteorological Coupled Spatial Transformer block, enabling the flexible coupling of AI-based meteorological forecasts while learning the interactions among pollutants and their diverse spatial responses. As for the lack of standardized datasets in air pollutants forecasting, we construct a comprehensive dataset covering 6 major pollutants across 75 cities in North China from 2018 to 2023, including ERA5 reanalysis data and FuXi-2.0 forecast data. Experimental results demonstrate that the proposed model outperforms state-of-the-art methods and validate the effectiveness of the proposed architecture.
☆ Incorporating Fairness Constraints into Archetypal Analysis
Archetypal Analysis (AA) is an unsupervised learning method that represents data as convex combinations of extreme patterns called archetypes. While AA provides interpretable and low-dimensional representations, it can inadvertently encode sensitive attributes, leading to fairness concerns. In this work, we propose Fair Archetypal Analysis (FairAA), a modified formulation that explicitly reduces the influence of sensitive group information in the learned projections. We also introduce FairKernelAA, a nonlinear extension that addresses fairness in more complex data distributions. Our approach incorporates a fairness regularization term while preserving the structure and interpretability of the archetypes. We evaluate FairAA and FairKernelAA on synthetic datasets, including linear, nonlinear, and multi-group scenarios, demonstrating their ability to reduce group separability -- as measured by mean maximum discrepancy and linear separability -- without substantially compromising explained variance. We further validate our methods on the real-world ANSUR I dataset, confirming their robustness and practical utility. The results show that FairAA achieves a favorable trade-off between utility and fairness, making it a promising tool for responsible representation learning in sensitive applications.
☆ DUSE: A Data Expansion Framework for Low-resource Automatic Modulation Recognition based on Active Learning
Although deep neural networks have made remarkable achievements in the field of automatic modulation recognition (AMR), these models often require a large amount of labeled data for training. However, in many practical scenarios, the available target domain data is scarce and difficult to meet the needs of model training. The most direct way is to collect data manually and perform expert annotation, but the high time and labor costs are unbearable. Another common method is data augmentation. Although it can enrich training samples to a certain extent, it does not introduce new data and therefore cannot fundamentally solve the problem of data scarcity. To address these challenges, we introduce a data expansion framework called Dynamic Uncertainty-driven Sample Expansion (DUSE). Specifically, DUSE uses an uncertainty scoring function to filter out useful samples from relevant AMR datasets and employs an active learning strategy to continuously refine the scorer. Extensive experiments demonstrate that DUSE consistently outperforms 8 coreset selection baselines in both class-balance and class-imbalance settings. Besides, DUSE exhibits strong cross-architecture generalization for unseen models.
☆ Expanding ML-Documentation Standards For Better Security
This article presents the current state of ML-security and of the documentation of ML-based systems, models and datasets in research and practice based on an extensive review of the existing literature. It shows a generally low awareness of security aspects among ML-practitioners and organizations and an often unstandardized approach to documentation, leading to overall low quality of ML-documentation. Existing standards are not regularly adopted in practice and IT-security aspects are often not included in documentation. Due to these factors, there is a clear need for improved security documentation in ML, as one step towards addressing the existing gaps in ML-security. To achieve this, we propose expanding existing documentation standards for ML-documentation to include a security section with specific security relevant information. Implementing this, a novel expanded method of documenting security requirements in ML-documentation is presented, based on the existing Model Cards and Datasheets for Datasets standards, but with the recommendation to adopt these findings in all ML-documentation.
comment: Accepted for publication at the 33rd IEEE International Requirements Engineering Workshop (REW 2025)
☆ Detecting In-Person Conversations in Noisy Real-World Environments with Smartwatch Audio and Motion Sensing
Social interactions play a crucial role in shaping human behavior, relationships, and societies. It encompasses various forms of communication, such as verbal conversation, non-verbal gestures, facial expressions, and body language. In this work, we develop a novel computational approach to detect a foundational aspect of human social interactions, in-person verbal conversations, by leveraging audio and inertial data captured with a commodity smartwatch in acoustically-challenging scenarios. To evaluate our approach, we conducted a lab study with 11 participants and a semi-naturalistic study with 24 participants. We analyzed machine learning and deep learning models with 3 different fusion methods, showing the advantages of fusing audio and inertial data to consider not only verbal cues but also non-verbal gestures in conversations. Furthermore, we perform a comprehensive set of evaluations across activities and sampling rates to demonstrate the benefits of multimodal sensing in specific contexts. Overall, our framework achieved 82.0$\pm$3.0% macro F1-score when detecting conversations in the lab and 77.2$\pm$1.8% in the semi-naturalistic setting.
☆ Can LLMs Find Fraudsters? Multi-level LLM Enhanced Graph Fraud Detection
Graph fraud detection has garnered significant attention as Graph Neural Networks (GNNs) have proven effective in modeling complex relationships within multimodal data. However, existing graph fraud detection methods typically use preprocessed node embeddings and predefined graph structures to reveal fraudsters, which ignore the rich semantic cues contained in raw textual information. Although Large Language Models (LLMs) exhibit powerful capabilities in processing textual information, it remains a significant challenge to perform multimodal fusion of processed textual embeddings with graph structures. In this paper, we propose a \textbf{M}ulti-level \textbf{L}LM \textbf{E}nhanced Graph Fraud \textbf{D}etection framework called MLED. In MLED, we utilize LLMs to extract external knowledge from textual information to enhance graph fraud detection methods. To integrate LLMs with graph structure information and enhance the ability to distinguish fraudsters, we design a multi-level LLM enhanced framework including type-level enhancer and relation-level enhancer. One is to enhance the difference between the fraudsters and the benign entities, the other is to enhance the importance of the fraudsters in different relations. The experiments on four real-world datasets show that MLED achieves state-of-the-art performance in graph fraud detection as a generalized framework that can be applied to existing methods.
Dataset-Adaptive Dimensionality Reduction IEEE VIS 2025
Selecting the appropriate dimensionality reduction (DR) technique and determining its optimal hyperparameter settings that maximize the accuracy of the output projections typically involves extensive trial and error, often resulting in unnecessary computational overhead. To address this challenge, we propose a dataset-adaptive approach to DR optimization guided by structural complexity metrics. These metrics quantify the intrinsic complexity of a dataset, predicting whether higher-dimensional spaces are necessary to represent it accurately. Since complex datasets are often inaccurately represented in two-dimensional projections, leveraging these metrics enables us to predict the maximum achievable accuracy of DR techniques for a given dataset, eliminating redundant trials in optimizing DR. We introduce the design and theoretical foundations of these structural complexity metrics. We quantitatively verify that our metrics effectively approximate the ground truth complexity of datasets and confirm their suitability for guiding dataset-adaptive DR workflow. Finally, we empirically show that our dataset-adaptive workflow significantly enhances the efficiency of DR optimization without compromising accuracy.
comment: IEEE VIS 2025 & IEEE Transactions on Visualization and Computer Graphics (TVCG)
☆ Recent results on searches with boosted Higgs bosons at CMS
The study of boosted Higgs bosons at the LHC provides a unique window to probe Higgs boson couplings at high energy scales and search for signs of physics beyond the standard model. In these proceedings, we present recent results on boosted Higgs boson searches at the CMS experiment, highlighting innovative reconstruction and tagging techniques that enhance sensitivity in this challenging regime.
comment: 6 pages, 3 figures, The Thirteenth Annual Large Hadron Collider Physics (LHCP2025)
☆ Online Training and Pruning of Deep Reinforcement Learning Networks
Scaling deep neural networks (NN) of reinforcement learning (RL) algorithms has been shown to enhance performance when feature extraction networks are used but the gained performance comes at the significant expense of increased computational and memory complexity. Neural network pruning methods have successfully addressed this challenge in supervised learning. However, their application to RL is underexplored. We propose an approach to integrate simultaneous training and pruning within advanced RL methods, in particular to RL algorithms enhanced by the Online Feature Extractor Network (OFENet). Our networks (XiNet) are trained to solve stochastic optimization problems over the RL networks' weights and the parameters of variational Bernoulli distributions for 0/1 Random Variables $\xi$ scaling each unit in the networks. The stochastic problem formulation induces regularization terms that promote convergence of the variational parameters to 0 when a unit contributes little to the performance. In this case, the corresponding structure is rendered permanently inactive and pruned from its network. We propose a cost-aware, sparsity-promoting regularization scheme, tailored to the DenseNet architecture of OFENets expressing the parameter complexity of involved networks in terms of the parameters of the RVs in these networks. Then, when matching this cost with the regularization terms, the many hyperparameters associated with them are automatically selected, effectively combining the RL objectives and network compression. We evaluate our method on continuous control benchmarks (MuJoCo) and the Soft Actor-Critic RL agent, demonstrating that OFENets can be pruned considerably with minimal loss in performance. Furthermore, our results confirm that pruning large networks during training produces more efficient and higher performing RL agents rather than training smaller networks from scratch.
comment: 25 pages, 5 figures, 4 tables
☆ d-DQIVAR: Data-centric Visual Analytics and Reasoning for Data Quality Improvement
Approaches to enhancing data quality (DQ) are classified into two main categories: data- and process-driven. However, prior research has predominantly utilized batch data preprocessing within the data-driven framework, which often proves insufficient for optimizing machine learning (ML) model performance and frequently leads to distortions in data characteristics. Existing studies have primarily focused on data preprocessing rather than genuine data quality improvement (DQI). In this paper, we introduce d-DQIVAR, a novel visual analytics system designed to facilitate DQI strategies aimed at improving ML model performance. Our system integrates visual analytics techniques that leverage both data-driven and process-driven approaches. Data-driven techniques tackle DQ issues such as imputation, outlier detection, deletion, format standardization, removal of duplicate records, and feature selection. Process-driven strategies encompass evaluating DQ and DQI procedures by considering DQ dimensions and ML model performance and applying the Kolmogorov-Smirnov test. We illustrate how our system empowers users to harness expert and domain knowledge effectively within a practical workflow through case studies, evaluations, and user studies.
☆ PoTPTQ: A Two-step Power-of-Two Post-training for LLMs ECAI 2025
Large Language Models (LLMs) have demonstrated remarkable performance across various natural language processing (NLP) tasks. However, their deployment is challenging due to the substantial computational resources required. Power-of-two (PoT) quantization is a general tool to counteract this difficulty. Albeit previous works on PoT quantization can be efficiently dequantized on CPUs using fixed-point addition, it showed less effectiveness on GPUs. The reason is entanglement of the sign bit and sequential bit manipulations needed for dequantization. We propose a novel POT quantization framework for LLM weights that (i) outperforms state-of-the-art accuracy in extremely low-precision number formats, and (ii) enables faster inference through more efficient dequantization. To maintain the accuracy of the quantized model, we introduce a two-step post-training algorithm: (i) initialize the quantization scales with a robust starting point, and (ii) refine these scales using a minimal calibration set. The performance of our PoT post-training algorithm surpasses the current state-of-the-art in integer quantization, particularly at low precisions such as 2- and 3-bit formats. Our PoT quantization accelerates the dequantization step required for the floating point inference and leads to $3.67\times$ speed up on a NVIDIA V100, and $1.63\times$ on a NVIDIA RTX 4090, compared to uniform integer dequantization.
comment: Accepted at ECAI 2025 (European Conference on Artificial Intelligence)
☆ The benefits of query-based KGQA systems for complex and temporal questions in LLM era
Large language models excel in question-answering (QA) yet still struggle with multi-hop reasoning and temporal questions. Query-based knowledge graph QA (KGQA) offers a modular alternative by generating executable queries instead of direct answers. We explore multi-stage query-based framework for WikiData QA, proposing multi-stage approach that enhances performance on challenging multi-hop and temporal benchmarks. Through generalization and rejection studies, we evaluate robustness across multi-hop and temporal QA datasets. Additionally, we introduce a novel entity linking and predicate matching method using CoT reasoning. Our results demonstrate the potential of query-based multi-stage KGQA framework for improving multi-hop and temporal QA with small language models. Code and data: https://github.com/ar2max/NLDB-KGQA-System
comment: 15 pages, 3 figures, 7 tables
☆ IAM: Efficient Inference through Attention Mapping between Different-scale LLMs ACL 2025
LLMs encounter significant challenges in resource consumption nowadays, especially with long contexts. Despite extensive efforts dedicate to enhancing inference efficiency, these methods primarily exploit internal sparsity within the models, without leveraging external information for optimization. We identify the high similarity of attention matrices across different-scale LLMs, which offers a novel perspective for optimization. We first conduct a comprehensive analysis of how to measure similarity, how to select mapping Layers and whether mapping is consistency. Based on these insights, we introduce the IAM framework, which achieves dual benefits of accelerated attention computation and reduced KV cache usage by performing attention mapping between small and large LLMs. Our experimental results demonstrate that IAM can accelerate prefill by 15% and reduce KV cache usage by 22.1% without appreciably sacrificing performance. Experiments on different series of models show the generalizability of IAM. Importantly, it is also orthogonal to many existing KV cache optimization methods, making it a versatile addition to the current toolkit for enhancing LLM efficiency.
comment: ACL 2025
☆ RNAMunin: A Deep Machine Learning Model for Non-coding RNA Discovery
Functional annotation of microbial genomes is often biased toward protein-coding genes, leaving a vast, unexplored landscape of non-coding RNAs (ncRNAs) that are critical for regulating bacterial and archaeal physiology, stress response and metabolism. Identifying ncRNAs directly from genomic sequence is a paramount challenge in bioinformatics and biology, essential for understanding the complete regulatory potential of an organism. This paper presents RNAMunin, a machine learning (ML) model that is capable of finding ncRNAs using genomic sequence alone. It is also computationally viable for large sequence datasets such as long read metagenomic assemblies with contigs totaling multiple Gbp. RNAMunin is trained on Rfam sequences extracted from approximately 60 Gbp of long read metagenomes from 16 San Francisco Estuary samples. We know of no other model that can detect ncRNAs based solely on genomic sequence at this scale. Since RNAMunin only requires genomic sequence as input, we do not need for an ncRNA to be transcribed to find it, i.e., we do not need transcriptomics data. We wrote this manuscript in a narrative style in order to best convey how RNAMunin was developed and how it works in detail. Unlike almost all current ML models, at approximately 1M parameters, RNAMunin is very small and very fast.
☆ Kevin: Multi-Turn RL for Generating CUDA Kernels
Writing GPU kernels is a challenging task and critical for AI systems' efficiency. It is also highly iterative: domain experts write code and improve performance through execution feedback. Moreover, it presents verifiable rewards like correctness and speedup, making it a natural environment to apply Reinforcement Learning (RL). To explicitly incorporate the iterative nature of this process into training, we develop a flexible multi-turn RL recipe that addresses unique challenges encountered in real-world settings, such as learning from long trajectories and effective reward attribution across turns. We present Kevin - K(ernel D)evin, the first model trained with multi-turn RL for CUDA kernel generation and optimization. In our evaluation setup, Kevin shows significant gains over its base model (QwQ-32B), improving correctness of generated kernels (in pure CUDA) from 56% to 82% and mean speedup from 0.53x to 1.10x of baseline (PyTorch Eager), and surpassing frontier models like o4-mini (0.78x). Finally, we study its behavior across test-time scaling axes: we found scaling serial refinement more beneficial than parallel sampling. In particular, when given more refinement turns, Kevin shows a higher rate of improvement.
☆ A Survey of Deep Learning for Geometry Problem Solving
Geometry problem solving is a key area of mathematical reasoning, which is widely involved in many important fields such as education, mathematical ability assessment of artificial intelligence, and multimodal ability assessment. In recent years, the rapid development of deep learning technology, especially the rise of multimodal large language models, has triggered a widespread research boom. This paper provides a survey of the applications of deep learning in geometry problem solving, including (i) a comprehensive summary of the relevant tasks in geometry problem solving; (ii) a thorough review of related deep learning methods; (iii) a detailed analysis of evaluation metrics and methods; and (iv) a critical discussion of the current challenges and future directions that can be explored. Our goal is to provide a comprehensive and practical reference of deep learning for geometry problem solving to promote further developments in this field. We create a continuously updated list of papers on GitHub: https://github.com/majianz/dl4gps.
comment: Work in progress
☆ Accelerating RF Power Amplifier Design via Intelligent Sampling and ML-Based Parameter Tuning
This paper presents a machine learning-accelerated optimization framework for RF power amplifier design that reduces simulation requirements by 65% while maintaining $\pm0.3$ to $\pm0.4$ dBm accuracy. The proposed method combines MaxMin Latin Hypercube Sampling with CatBoost gradient boosting to intelligently explore multidimensional parameter spaces. Instead of exhaustively simulating all parameter combinations to achieve target P2dB compression specifications, our approach strategically selects approximately 35% of critical simulation points. The framework processes ADS netlists, executes harmonic balance simulations on the reduced dataset, and trains a CatBoost model to predict P2dB performance across the entire design space. Validation across 15 PA operating modes yields an average $R^2$ of 0.901, with the system ranking parameter combinations by their likelihood of meeting target specifications. The integrated solution delivers 58.24% to 77.78% reduction in simulation time through automated GUI-based workflows, enabling rapid design iterations without compromising accuracy standards required for production RF circuits.
comment: This paper is a pre-print version and has been submitted to the IEEE International Conference on Future Machine Learning and Data Science (FMLDS 2025)
☆ From Generative to Episodic: Sample-Efficient Replicable Reinforcement Learning
The epidemic failure of replicability across empirical science and machine learning has recently motivated the formal study of replicable learning algorithms [Impagliazzo et al. (2022)]. In batch settings where data comes from a fixed i.i.d. source (e.g., hypothesis testing, supervised learning), the design of data-efficient replicable algorithms is now more or less understood. In contrast, there remain significant gaps in our knowledge for control settings like reinforcement learning where an agent must interact directly with a shifting environment. Karbasi et. al show that with access to a generative model of an environment with $S$ states and $A$ actions (the RL 'batch setting'), replicably learning a near-optimal policy costs only $\tilde{O}(S^2A^2)$ samples. On the other hand, the best upper bound without a generative model jumps to $\tilde{O}(S^7 A^7)$ [Eaton et al. (2024)] due to the substantial difficulty of environment exploration. This gap raises a key question in the broader theory of replicability: Is replicable exploration inherently more expensive than batch learning? Is sample-efficient replicable RL even possible? In this work, we (nearly) resolve this problem (for low-horizon tabular MDPs): exploration is not a significant barrier to replicable learning! Our main result is a replicable RL algorithm on $\tilde{O}(S^2A)$ samples, bridging the gap between the generative and episodic settings. We complement this with a matching $\tilde{\Omega}(S^2A)$ lower bound in the generative setting (under the common parallel sampling assumption) and an unconditional lower bound in the episodic setting of $\tilde{\Omega}(S^2)$ showcasing the near-optimality of our algorithm with respect to the state space $S$.
comment: 67 pages
☆ AFPM: Alignment-based Frame Patch Modeling for Cross-Dataset EEG Decoding
Electroencephalogram (EEG) decoding models for brain-computer interfaces (BCIs) struggle with cross-dataset learning and generalization due to channel layout inconsistencies, non-stationary signal distributions, and limited neurophysiological prior integration. To address these issues, we propose a plug-and-play Alignment-Based Frame-Patch Modeling (AFPM) framework, which has two main components: 1) Spatial Alignment, which selects task-relevant channels based on brain-region priors, aligns EEG distributions across domains, and remaps the selected channels to a unified layout; and, 2) Frame-Patch Encoding, which models multi-dataset signals into unified spatiotemporal patches for EEG decoding. Compared to 17 state-of-the-art approaches that need dataset-specific tuning, the proposed calibration-free AFPM achieves performance gains of up to 4.40% on motor imagery and 3.58% on event-related potential tasks. To our knowledge, this is the first calibration-free cross-dataset EEG decoding framework, substantially enhancing the practicalness of BCIs in real-world applications.
☆ Resampling strategies for imbalanced regression: a survey and empirical analysis
Imbalanced problems can arise in different real-world situations, and to address this, certain strategies in the form of resampling or balancing algorithms are proposed. This issue has largely been studied in the context of classification, and yet, the same problem features in regression tasks, where target values are continuous. This work presents an extensive experimental study comprising various balancing and predictive models, and wich uses metrics to capture important elements for the user and to evaluate the predictive model in an imbalanced regression data context. It also proposes a taxonomy for imbalanced regression approaches based on three crucial criteria: regression model, learning process, and evaluation metrics. The study offers new insights into the use of such strategies, highlighting the advantages they bring to each model's learning process, and indicating directions for further studies. The code, data and further information related to the experiments performed herein can be found on GitHub: https://github.com/JusciAvelino/imbalancedRegression.
☆ Imbalanced Regression Pipeline Recommendation
Imbalanced problems are prevalent in various real-world scenarios and are extensively explored in classification tasks. However, they also present challenges for regression tasks due to the rarity of certain target values. A common alternative is to employ balancing algorithms in preprocessing to address dataset imbalance. However, due to the variety of resampling methods and learning models, determining the optimal solution requires testing many combinations. Furthermore, the learning model, dataset, and evaluation metric affect the best strategies. This work proposes the Meta-learning for Imbalanced Regression (Meta-IR) framework, which diverges from existing literature by training meta-classifiers to recommend the best pipeline composed of the resampling strategy and learning model per task in a zero-shot fashion. The meta-classifiers are trained using a set of meta-features to learn how to map the meta-features to the classes indicating the best pipeline. We propose two formulations: Independent and Chained. Independent trains the meta-classifiers to separately indicate the best learning algorithm and resampling strategy. Chained involves a sequential procedure where the output of one meta-classifier is used as input for another to model intrinsic relationship factors. The Chained scenario showed superior performance, suggesting a relationship between the learning algorithm and the resampling strategy per task. Compared with AutoML frameworks, Meta-IR obtained better results. Moreover, compared with baselines of six learning algorithms and six resampling algorithms plus no resampling, totaling 42 (6 X 7) configurations, Meta-IR outperformed all of them. The code, data, and further information of the experiments can be found on GitHub: https://github.com/JusciAvelino/Meta-IR.
☆ Newfluence: Boosting Model interpretability and Understanding in High Dimensions
The increasing complexity of machine learning (ML) and artificial intelligence (AI) models has created a pressing need for tools that help scientists, engineers, and policymakers interpret and refine model decisions and predictions. Influence functions, originating from robust statistics, have emerged as a popular approach for this purpose. However, the heuristic foundations of influence functions rely on low-dimensional assumptions where the number of parameters $p$ is much smaller than the number of observations $n$. In contrast, modern AI models often operate in high-dimensional regimes with large $p$, challenging these assumptions. In this paper, we examine the accuracy of influence functions in high-dimensional settings. Our theoretical and empirical analyses reveal that influence functions cannot reliably fulfill their intended purpose. We then introduce an alternative approximation, called Newfluence, that maintains similar computational efficiency while offering significantly improved accuracy. Newfluence is expected to provide more accurate insights than many existing methods for interpreting complex AI models and diagnosing their issues. Moreover, the high-dimensional framework we develop in this paper can also be applied to analyze other popular techniques, such as Shapley values.
☆ Choosing the Better Bandit Algorithm under Data Sharing: When Do A/B Experiments Work?
We study A/B experiments that are designed to compare the performance of two recommendation algorithms. Prior work has shown that the standard difference-in-means estimator is biased in estimating the global treatment effect (GTE) due to a particular form of interference between experimental units. Specifically, units under the treatment and control algorithms contribute to a shared pool of data that subsequently train both algorithms, resulting in interference between the two groups. The bias arising from this type of data sharing is known as "symbiosis bias". In this paper, we highlight that, for decision-making purposes, the sign of the GTE often matters more than its precise magnitude when selecting the better algorithm. We formalize this insight under a multi-armed bandit framework and theoretically characterize when the sign of the expected GTE estimate under data sharing aligns with or contradicts the sign of the true GTE. Our analysis identifies the level of exploration versus exploitation as a key determinant of how symbiosis bias impacts algorithm selection.
☆ A Policy-Improved Deep Deterministic Policy Gradient Framework for the Discount Order Acceptance Strategy of Ride-hailing Drivers
The rapid expansion of platform integration has emerged as an effective solution to mitigate market fragmentation by consolidating multiple ride-hailing platforms into a single application. To address heterogeneous passenger preferences, third-party integrators provide Discount Express service delivered by express drivers at lower trip fares. For the individual platform, encouraging broader participation of drivers in Discount Express services has the potential to expand the accessible demand pool and improve matching efficiency, but often at the cost of reduced profit margins. This study aims to dynamically manage drivers' acceptance of Discount Express from the perspective of individual platforms. The lack of historical data under the new business model necessitates online learning. However, early-stage exploration through trial and error can be costly in practice, highlighting the need for reliable early-stage performance in real-world deployment. To address these challenges, this study formulates the decision regarding the proportion of drivers' acceptance behavior as a continuous control task. In response to the high stochasticity, the opaque matching mechanisms employed by third-party integrator, and the limited availability of historical data, we propose a policy-improved deep deterministic policy gradient (pi-DDPG) framework. The proposed framework incorporates a refiner module to boost policy performance during the early training phase, leverages a convolutional long short-term memory network to effectively capture complex spatiotemporal patterns, and adopts a prioritized experience replay mechanism to enhance learning efficiency. A simulator based on a real-world dataset is developed to validate the effectiveness of the proposed pi-DDPG. Numerical experiments demonstrate that pi-DDPG achieves superior learning efficiency and significantly reduces early-stage training losses.
☆ OrdShap: Feature Position Importance for Sequential Black-Box Models
Sequential deep learning models excel in domains with temporal or sequential dependencies, but their complexity necessitates post-hoc feature attribution methods for understanding their predictions. While existing techniques quantify feature importance, they inherently assume fixed feature ordering - conflating the effects of (1) feature values and (2) their positions within input sequences. To address this gap, we introduce OrdShap, a novel attribution method that disentangles these effects by quantifying how a model's predictions change in response to permuting feature position. We establish a game-theoretic connection between OrdShap and Sanchez-Berganti\~nos values, providing a theoretically grounded approach to position-sensitive attribution. Empirical results from health, natural language, and synthetic datasets highlight OrdShap's effectiveness in capturing feature value and feature position attributions, and provide deeper insight into model behavior.
♻ ☆ Diffused Responsibility: Analyzing the Energy Consumption of Generative Text-to-Audio Diffusion Models SP
Text-to-audio models have recently emerged as a powerful technology for generating sound from textual descriptions. However, their high computational demands raise concerns about energy consumption and environmental impact. In this paper, we conduct an analysis of the energy usage of 7 state-of-the-art text-to-audio diffusion-based generative models, evaluating to what extent variations in generation parameters affect energy consumption at inference time. We also aim to identify an optimal balance between audio quality and energy consumption by considering Pareto-optimal solutions across all selected models. Our findings provide insights into the trade-offs between performance and environmental impact, contributing to the development of more efficient generative audio models.
comment: Accepted at WASPAA 2025
♻ ☆ MARS: Unleashing the Power of Variance Reduction for Training Large Models
Training deep neural networks--and more recently, large models demands efficient and scalable optimizers. Adaptive gradient algorithms like Adam, AdamW, and their variants have been central to this task. Despite the development of numerous variance reduction algorithms in the past decade aimed at accelerating stochastic optimization in both convex and nonconvex settings, variance reduction has not found widespread success in training deep neural networks or large language models. Consequently, it has remained a less favored approach in modern AI. In this paper, to unleash the power of variance reduction for efficient training of large models, we propose a unified optimization framework, MARS (Make vAriance Reduction Shine), which reconciles preconditioned gradient methods with variance reduction via a scaled stochastic recursive momentum technique. Within our framework, we introduce three instances of MARS that leverage preconditioned gradient updates based on AdamW, Lion, and Shampoo, respectively. We also draw a connection between our algorithms and existing optimizers. Experimental results on training GPT-2 models indicate that MARS consistently outperforms AdamW by a large margin. The implementation of MARS is available at https://github.com/AGI-Arena/MARS.
comment: 35 pages, 19 figures, 12 tables
♻ ☆ Navigating the Social Welfare Frontier: Portfolios for Multi-objective Reinforcement Learning
In many real-world applications of reinforcement learning (RL), deployed policies have varied impacts on different stakeholders, creating challenges in reaching consensus on how to effectively aggregate their preferences. Generalized $p$-means form a widely used class of social welfare functions for this purpose, with broad applications in fair resource allocation, AI alignment, and decision-making. This class includes well-known welfare functions such as Egalitarian, Nash, and Utilitarian welfare. However, selecting the appropriate social welfare function is challenging for decision-makers, as the structure and outcomes of optimal policies can be highly sensitive to the choice of $p$. To address this challenge, we study the concept of an $\alpha$-approximate portfolio in RL, a set of policies that are approximately optimal across the family of generalized $p$-means for all $p \in [-\infty, 1]$. We propose algorithms to compute such portfolios and provide theoretical guarantees on the trade-offs among approximation factor, portfolio size, and computational efficiency. Experimental results on synthetic and real-world datasets demonstrate the effectiveness of our approach in summarizing the policy space induced by varying $p$ values, empowering decision-makers to navigate this landscape more effectively.
♻ ☆ The Utility of the Virtual Imaging Trials Methodology for Objective Characterization of AI Systems and Training Data
Purpose: The credibility of Artificial Intelligence (AI) models for medical imaging continues to be a challenge, affected by the diversity of models, the data used to train the models, and applicability of their combination to produce reproducible results for new data. Approach: In this work we aimed to explore if the emerging Virtual Imaging Trials (VIT) methodologies can provide an objective resource to approach this challenge. The study was conducted for the case example of COVID-19 diagnosis using clinical and virtual computed tomography (CT) and chest radiography (CXR) processed with convolutional neural networks. Multiple AI models were developed and tested using 3D ResNet-like and 2D EfficientNetv2 architectures across diverse datasets. Results: The performance differences were evaluated in terms of the area under the curve (AUC) and the DeLong method for AUC confidence intervals. The models trained on the most diverse datasets showed the highest external testing performance, with AUC values ranging from 0.73-0.76 for CT and 0.70-0.73 for CXR. Internal testing yielded higher AUC values (0.77 -0.85 for CT and 0.77-1.0 for CXR), highlighting a substantial drop in performance during external validation, which underscores the importance of diverse and comprehensive training and testing data. Most notably, VIT approach provided objective assessment of the utility of diverse models and datasets while further providing insight into the influence of dataset characteristics, patient factors, and imaging physics on AI efficacy. Conclusions: The VIT approach can be used to enhance model transparency and reliability, offering nuanced insights into the factors driving AI performance and bridging the gap between experimental and clinical settings.
comment: 8 figures, 4 Tables
♻ ☆ PBM-VFL: Vertical Federated Learning with Feature and Sample Privacy
We present Poisson Binomial Mechanism Vertical Federated Learning (PBM-VFL), a communication-efficient Vertical Federated Learning algorithm with Differential Privacy guarantees. PBM-VFL combines Secure Multi-Party Computation with the recently introduced Poisson Binomial Mechanism to protect parties' private datasets during model training. We define the novel concept of feature privacy and analyze end-to-end feature and sample privacy of our algorithm. We compare sample privacy loss in VFL with privacy loss in HFL. We also provide the first theoretical characterization of the relationship between privacy budget, convergence error, and communication cost in differentially-private VFL. Finally, we empirically show that our model performs well with high levels of privacy.
♻ ☆ DUNIA: Pixel-Sized Embeddings via Cross-Modal Alignment for Earth Observation Applications
Significant efforts have been directed towards adapting self-supervised multimodal learning for Earth observation applications. However, most current methods produce coarse patch-sized embeddings, limiting their effectiveness and integration with other modalities like LiDAR. To close this gap, we present DUNIA, an approach to learn pixel-sized embeddings through cross-modal alignment between images and full-waveform LiDAR data. As the model is trained in a contrastive manner, the embeddings can be directly leveraged in the context of a variety of environmental monitoring tasks in a zero-shot setting. In our experiments, we demonstrate the effectiveness of the embeddings for seven such tasks: canopy height mapping, fractional canopy cover, land cover mapping, tree species identification, plant area index, crop type classification, and per-pixel waveform-based vertical structure mapping. The results show that the embeddings, along with zero-shot classifiers, often outperform specialized supervised models, even in low-data regimes. In the fine-tuning setting, we show strong performances near or better than the state-of-the-art on five out of six tasks.
comment: 26 pages, 8 figures
♻ ☆ MirrorCBO: A consensus-based optimization method in the spirit of mirror descent
In this work we propose MirrorCBO, a consensus-based optimization (CBO) method which generalizes standard CBO in the same way that mirror descent generalizes gradient descent. For this we apply the CBO methodology to a swarm of dual particles and retain the primal particle positions by applying the inverse of the mirror map, which we parametrize as the subdifferential of a strongly convex function $\phi$. In this way, we combine the advantages of a derivative-free non-convex optimization algorithm with those of mirror descent. As a special case, the method extends CBO to optimization problems with convex constraints. Assuming bounds on the Bregman distance associated to $\phi$, we provide asymptotic convergence results for MirrorCBO with explicit exponential rate. Another key contribution is an exploratory numerical study of this new algorithm across different application settings, focusing on (i) sparsity-inducing optimization, and (ii) constrained optimization, demonstrating the competitive performance of MirrorCBO. We observe empirically that the method can also be used for optimization on (non-convex) submanifolds of Euclidean space, can be adapted to mirrored versions of other recent CBO variants, and that it inherits from mirror descent the capability to select desirable minimizers, like sparse ones. We also include an overview of recent CBO approaches for constrained optimization and compare their performance to MirrorCBO.
comment: 66 pages, 18 figures, 19 tables
♻ ☆ Simple Mechanistic Explanations for Out-Of-Context Reasoning ICML 2025
Out-of-context reasoning (OOCR) is a phenomenon in which fine-tuned LLMs exhibit surprisingly deep out-of-distribution generalization. Rather than learning shallow heuristics, they implicitly internalize and act on the consequences of observations scattered throughout the fine-tuning data. In this work, we investigate this phenomenon mechanistically and find that many instances of OOCR in the literature have a simple explanation: the LoRA fine-tuning essentially adds a constant steering vector, steering the model towards a general concept. This improves performance on the fine-tuning task and in many other concept-related domains, causing the surprising generalization. Moreover, we can directly train steering vectors for these tasks from scratch, which also induces OOCR. We find that our results hold even for a task that seems like it must involve conditional behavior (model backdoors); it turns out that unconditionally adding a steering vector is sufficient. Overall, our work presents one explanation of what gets learned during fine-tuning for OOCR tasks, contributing to the key question of why LLMs can reason out of context, an advanced capability that is highly relevant to their safe and reliable deployment.
comment: ICML 2025 Workshop R2-FM
♻ ☆ Large Language Models are Unreliable for Cyber Threat Intelligence
Several recent works have argued that Large Language Models (LLMs) can be used to tame the data deluge in the cybersecurity field, by improving the automation of Cyber Threat Intelligence (CTI) tasks. This work presents an evaluation methodology that other than allowing to test LLMs on CTI tasks when using zero-shot learning, few-shot learning and fine-tuning, also allows to quantify their consistency and their confidence level. We run experiments with three state-of-the-art LLMs and a dataset of 350 threat intelligence reports and present new evidence of potential security risks in relying on LLMs for CTI. We show how LLMs cannot guarantee sufficient performance on real-size reports while also being inconsistent and overconfident. Few-shot learning and fine-tuning only partially improve the results, thus posing doubts about the possibility of using LLMs for CTI scenarios, where labelled datasets are lacking and where confidence is a fundamental factor.
♻ ☆ BondMatcher: H-Bond Stability Analysis in Molecular Systems IEEE VIS 2025
This application paper investigates the stability of hydrogen bonds (H-bonds), as characterized by the Quantum Theory of Atoms in Molecules (QTAIM). First, we contribute a database of 4544 electron densities associated to four isomers of water hexamers (the so-called Ring, Book, Cage and Prism), generated by distorting their equilibrium geometry under various structural perturbations, modeling the natural dynamic behavior of molecular systems. Second, we present a new stability measure, called bond occurrence rate, associating each bond path present at equilibrium with its rate of occurrence within the input ensemble. We also provide an algorithm, called BondMatcher, for its automatic computation, based on a tailored, geometry-aware partial isomorphism estimation between the extremum graphs of the considered electron densities. Our new stability measure allows for the automatic identification of densities lacking H-bond paths, enabling further visual inspections. Specifically, the topological analysis enabled by our framework corroborates experimental observations and provides refined geometrical criteria for characterizing the disappearance of H-bond paths. Our electron density database and our C++ implementation are available at this address: https://github.com/thom-dani/BondMatcher.
comment: To appear in IEEE Transactions on Visualization and Computer Graphics (Proc. of IEEE VIS 2025)
♻ ☆ Towards Understanding Link Predictor Generalizability Under Distribution Shifts KDD 25
State-of-the-art link prediction (LP) models demonstrate impressive benchmark results. However, popular benchmark datasets often assume that training, validation, and testing samples are representative of the overall dataset distribution. In real-world situations, this assumption is often incorrect; uncontrolled factors lead new dataset samples to come from a different distribution than training samples. Additionally, the majority of recent work with graph dataset shift focuses on node- and graph-level tasks, largely ignoring link-level tasks. To bridge this gap, we introduce a novel splitting strategy, known as LPShift, which utilizes structural properties to induce a controlled distribution shift. We verify LPShift's effect through empirical evaluation of SOTA LP models on 16 LPShift variants of original dataset splits, with results indicating drastic changes to model performance. Additional experiments demonstrate graph structure has a strong influence on the success of current generalization methods. Source Code Available Here: https://github.com/revolins/LPShift
comment: KDD 25' Datasets & Benchmarks Track, 14 pages, 8 figures, 18 tables
♻ ☆ Exploring and Analyzing Wildland Fire Data Via Machine Learning Techniques
This research project investigated the correlation between a 10 Hz time series of thermocouple temperatures and turbulent kinetic energy (TKE) computed from wind speeds collected from a small experimental prescribed burn at the Silas Little Experimental Forest in New Jersey, USA. The primary objective of this project was to explore the potential for using thermocouple temperatures as predictors for estimating the TKE produced by a wildland fire. Machine learning models, including Deep Neural Networks, Random Forest Regressor, Gradient Boosting, and Gaussian Process Regressor, are employed to assess the potential for thermocouple temperature perturbations to predict TKE values. Data visualization and correlation analyses reveal patterns and relationships between thermocouple temperatures and TKE, providing insight into the underlying dynamics. The project achieves high accuracy in predicting TKE by employing various machine learning models despite a weak correlation between the predictors and the target variable. The results demonstrate significant success, particularly from regression models, in accurately estimating the TKE. The research findings contribute to fire behavior and smoke modeling science, emphasizing the importance of incorporating machine learning approaches and identifying complex relationships between fine-scale fire behavior and turbulence. Accurate TKE estimation using thermocouple temperatures allows for the refinement of models that can inform decision-making in fire management strategies, facilitate effective risk mitigation, and optimize fire management efforts. This project highlights the valuable role of machine learning techniques in analyzing wildland fire data, showcasing their potential to advance fire research and management practices.
comment: This version has been significantly updated and superseded by a more complete and revised version under the new title "Leveraging Advanced Machine Learning to Predict Turbulence Dynamics from Temperature Observations at an Experimental Prescribed Fire", available at arXiv ID: arXiv:2507.11012
♻ ☆ Distilling Invariant Representations with Dual Augmentation
Knowledge distillation (KD) has been widely used to transfer knowledge from large, accurate models (teachers) to smaller, efficient ones (students). Recent methods have explored enforcing consistency by incorporating causal interpretations to distill invariant representations. In this work, we extend this line of research by introducing a dual augmentation strategy to promote invariant feature learning in both teacher and student models. Our approach leverages different augmentations applied to both models during distillation, pushing the student to capture robust, transferable features. This dual augmentation strategy complements invariant causal distillation by ensuring that the learned representations remain stable across a wider range of data variations and transformations. Extensive experiments on CIFAR-100 demonstrate the effectiveness of this approach, achieving competitive results in same-architecture KD.
comment: After further review, we determined that the submission does not meet the quality standards we intended
♻ ☆ Bridging Predictive Coding and MDL: A Two-Part Code Framework for Deep Learning
We present the first theoretical framework that connects predictive coding (PC), a biologically inspired local learning rule, with the minimum description length (MDL) principle in deep networks. We prove that layerwise PC performs block-coordinate descent on the MDL two-part code objective, thereby jointly minimizing empirical risk and model complexity. Using Hoeffding's inequality and a prefix-code prior, we derive a novel generalization bound of the form $R(\theta) \le \hat{R}(\theta) + \frac{L(\theta)}{N}$, capturing the tradeoff between fit and compression. We further prove that each PC sweep monotonically decreases the empirical two-part codelength, yielding tighter high-probability risk bounds than unconstrained gradient descent. Finally, we show that repeated PC updates converge to a block-coordinate stationary point, providing an approximate MDL-optimal solution. To our knowledge, this is the first result offering formal generalization and convergence guarantees for PC-trained deep models, positioning PC as a theoretically grounded and biologically plausible alternative to backpropagation.
comment: 24 pages, 2 figures
♻ ☆ Planning-Aware Code Infilling via Horizon-Length Prediction
Fill-in-the-Middle (FIM), or infilling, has become integral to code language models, enabling generation of missing code given both left and right contexts. However, the current FIM training paradigm which performs next-token prediction (NTP) over reordered sequence often leads to models struggling to generate content that aligns well with the surrounding context. We hypothesize that NTP alone is insufficient for models to learn effective planning conditioned on the distant right context, a critical factor for successful code infilling. To overcome this, we propose Horizon-Length Prediction (HLP), a novel training objective that teaches models to predict the number of remaining middle tokens at each step. HLP advances FIM with lookahead planning, enabling models to inherently learn infilling boundaries for arbitrary left and right contexts without relying on dataset-specific post-processing. Our evaluation across different model families and sizes shows that HLP significantly improves FIM performance by up to 24% relatively on diverse benchmarks, across file-level and repository-level. Furthermore, the enhanced planning capability gained through HLP boosts model performance on code reasoning. Importantly, HLP incurs negligible training overhead and no additional inference cost, ensuring its practicality for real-world scenarios.
♻ ☆ Sparse Orthogonal Parameters Tuning for Continual Learning
Continual learning methods based on pre-trained models (PTM) have recently gained attention which adapt to successive downstream tasks without catastrophic forgetting. These methods typically refrain from updating the pre-trained parameters and instead employ additional adapters, prompts, and classifiers. In this paper, we from a novel perspective investigate the benefit of sparse orthogonal parameters for continual learning. We found that merging sparse orthogonality of models learned from multiple streaming tasks has great potential in addressing catastrophic forgetting. Leveraging this insight, we propose a novel yet effective method called SoTU (Sparse Orthogonal Parameters TUning). We hypothesize that the effectiveness of SoTU lies in the transformation of knowledge learned from multiple domains into the fusion of orthogonal delta parameters. Experimental evaluations on diverse CL benchmarks demonstrate the effectiveness of the proposed approach. Notably, SoTU achieves optimal feature representation for streaming data without necessitating complex classifier designs, making it a Plug-and-Play solution.
♻ ☆ Active Deep Kernel Learning of Molecular Properties: Realizing Dynamic Structural Embeddings
As vast databases of chemical identities become increasingly available, the challenge shifts to how we effectively explore and leverage these resources to study molecular properties. This paper presents an active learning approach for molecular discovery using Deep Kernel Learning (DKL), demonstrated on the QM9 dataset. DKL links structural embeddings directly to properties, creating organized latent spaces that prioritize relevant property information. By iteratively recalculating embedding vectors in alignment with target properties, DKL uncovers concentrated maxima representing key molecular properties and reveals unexplored regions with potential for innovation. This approach underscores DKL's potential in advancing molecular research and discovery.
♻ ☆ GHPO: Adaptive Guidance for Stable and Efficient LLM Reinforcement Learning
Reinforcement Learning with Verifiable Rewards (RLVR) has recently emerged as a powerful paradigm for facilitating the self-improvement of large language models (LLMs), particularly in the domain of complex reasoning tasks. However, prevailing on-policy RL methods often contend with significant training instability and inefficiency. This is primarily due to a capacity-difficulty mismatch, where the complexity of training data frequently outpaces the model's current capabilities, leading to critically sparse reward signals and stalled learning progress. This challenge is particularly acute for smaller, more resource-efficient LLMs. To overcome this, we introduce the Guided Hybrid Policy Optimization (GHPO), a novel difficulty-aware reinforcement learning framework. GHPO dynamically calibrates task difficulty by employing adaptive prompt refinement to provide targeted guidance. This unique approach adaptively balances direct imitation learning for problems currently beyond the model's reach with exploration-based reinforcement learning for more manageable tasks, effectively creating a smooth and optimized learning curriculum. Extensive experiments demonstrate that GHPO achieves an average performance gain of approximately 5% across six challenging mathematics benchmarks, consistently outperforming strong on-policy reinforcement learning and curriculum learning baselines. Further analysis confirms that our framework significantly enhances both training stability and final reasoning performance, thus offering a scalable and efficient solution for developing powerful and robust reasoning models.
comment: Code avaiable at https://github.com/hkgc-1/GHPO
♻ ☆ Quantifying calibration error in modern neural networks through evidence based theory
Trustworthiness in neural networks is crucial for their deployment in critical applications, where reliability, confidence, and uncertainty play pivotal roles in decision-making. Traditional performance metrics such as accuracy and precision fail to capture these aspects, particularly in cases where models exhibit overconfidence. To address these limitations, this paper introduces a novel framework for quantifying the trustworthiness of neural networks by incorporating subjective logic into the evaluation of Expected Calibration Error (ECE). This method provides a comprehensive measure of trust, disbelief, and uncertainty by clustering predicted probabilities and fusing opinions using appropriate fusion operators. We demonstrate the effectiveness of this approach through experiments on MNIST and CIFAR-10 datasets, where post-calibration results indicate improved trustworthiness. The proposed framework offers a more interpretable and nuanced assessment of AI models, with potential applications in sensitive domains such as healthcare and autonomous systems.
comment: Accepted at FUSION 2025 Conference
♻ ☆ MVP-Shapley: Feature-based Modeling for Evaluating the Most Valuable Player in Basketball
The burgeoning growth of the esports and multiplayer online gaming community has highlighted the critical importance of evaluating the Most Valuable Player (MVP). The establishment of an explainable and practical MVP evaluation method is very challenging. In our study, we specifically focus on play-by-play data, which records related events during the game, such as assists and points. We aim to address the challenges by introducing a new MVP evaluation framework, denoted as \oursys, which leverages Shapley values. This approach encompasses feature processing, win-loss model training, Shapley value allocation, and MVP ranking determination based on players' contributions. Additionally, we optimize our algorithm to align with expert voting results from the perspective of causality. Finally, we substantiated the efficacy of our method through validation using the NBA dataset and the Dunk City Dynasty dataset and implemented online deployment in the industry.
♻ ☆ AnnoPage Dataset: Dataset of Non-Textual Elements in Documents with Fine-Grained Categorization ICDAR2025
We introduce the AnnoPage Dataset, a novel collection of 7,550 pages from historical documents, primarily in Czech and German, spanning from 1485 to the present, focusing on the late 19th and early 20th centuries. The dataset is designed to support research in document layout analysis and object detection. Each page is annotated with axis-aligned bounding boxes (AABB) representing elements of 25 categories of non-textual elements, such as images, maps, decorative elements, or charts, following the Czech Methodology of image document processing. The annotations were created by expert librarians to ensure accuracy and consistency. The dataset also incorporates pages from multiple, mainly historical, document datasets to enhance variability and maintain continuity. The dataset is divided into development and test subsets, with the test set carefully selected to maintain the category distribution. We provide baseline results using YOLO and DETR object detectors, offering a reference point for future research. The AnnoPage Dataset is publicly available on Zenodo (https://doi.org/10.5281/zenodo.12788419), along with ground-truth annotations in YOLO format.
comment: 17 pages, 2 tables, 7 figures; Accepted to GREC Workshop at ICDAR2025
♻ ☆ ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy ICML 2025
Large-scale cell microscopy screens are used in drug discovery and molecular biology research to study the effects of millions of chemical and genetic perturbations on cells. To use these images in downstream analysis, we need models that can map each image into a feature space that represents diverse biological phenotypes consistently, in the sense that perturbations with similar biological effects have similar representations. In this work, we present the largest foundation model for cell microscopy data to date, a new 1.9 billion-parameter ViT-G/8 MAE trained on over 8 billion microscopy image crops. Compared to a previous published ViT-L/8 MAE, our new model achieves a 60% improvement in linear separability of genetic perturbations and obtains the best overall performance on whole-genome biological relationship recall and replicate consistency benchmarks. Beyond scaling, we developed two key methods that improve performance: (1) training on a curated and diverse dataset; and, (2) using biologically motivated linear probing tasks to search across each transformer block for the best candidate representation of whole-genome screens. We find that many self-supervised vision transformers, pretrained on either natural or microscopy images, yield significantly more biologically meaningful representations of microscopy images in their intermediate blocks than in their typically used final blocks. More broadly, our approach and results provide insights toward a general strategy for successfully building foundation models for large-scale biological data.
comment: ICML 2025 main-track paper (42nd International Conference on Machine Learning). Formerly appeared as best paper runner-up at NeurIPS 2024 Foundation Models for Science Workshop (38th Conference on Neural Information Processing Systems). 18 pages, 7 figures
♻ ☆ On the Statistical Properties of Generative Adversarial Models for Low Intrinsic Data Dimension
Despite the remarkable empirical successes of Generative Adversarial Networks (GANs), the theoretical guarantees for their statistical accuracy remain rather pessimistic. In particular, the data distributions on which GANs are applied, such as natural images, are often hypothesized to have an intrinsic low-dimensional structure in a typically high-dimensional feature space, but this is often not reflected in the derived rates in the state-of-the-art analyses. In this paper, we attempt to bridge the gap between the theory and practice of GANs and their bidirectional variant, Bi-directional GANs (BiGANs), by deriving statistical guarantees on the estimated densities in terms of the intrinsic dimension of the data and the latent space. We analytically show that if one has access to $n$ samples from the unknown target distribution and the network architectures are properly chosen, the expected Wasserstein-1 distance of the estimates from the target scales as $O\left( n^{-1/d_\mu } \right)$ for GANs and $\tilde{O}\left( n^{-1/(d_\mu+\ell)} \right)$ for BiGANs, where $d_\mu$ and $\ell$ are the upper Wasserstein-1 dimension of the data-distribution and latent-space dimension, respectively. The theoretical analyses not only suggest that these methods successfully avoid the curse of dimensionality, in the sense that the exponent of $n$ in the error rates does not depend on the data dimension but also serve to bridge the gap between the theoretical analyses of GANs and the known sharp rates from optimal transport literature. Additionally, we demonstrate that GANs can effectively achieve the minimax optimal rate even for non-smooth underlying distributions, with the use of interpolating generator networks.
comment: Journal of Machine Learning Research (2025), volume 26
♻ ☆ RACER: Rational Artificial Intelligence Car-following-model Enhanced by Reality
This paper introduces RACER, the Rational Artificial Intelligence Car-following model Enhanced by Reality, a cutting-edge deep learning car-following model, that satisfies partial derivative constraints, designed to predict Adaptive Cruise Control (ACC) driving behavior while staying theoretically feasible. Unlike conventional models, RACER effectively integrates Rational Driving Constraints (RDCs), crucial tenets of actual driving, resulting in strikingly accurate and realistic predictions. Against established models like the Optimal Velocity Relative Velocity (OVRV), a car-following Neural Network (NN), and a car-following Physics-Informed Neural Network (PINN), RACER excels across key metrics, such as acceleration, velocity, and spacing. Notably, it displays a perfect adherence to the RDCs, registering zero violations, in stark contrast to other models. This study highlights the immense value of incorporating physical constraints within AI models, especially for augmenting safety measures in transportation. It also paves the way for future research to test these models against human driving data, with the potential to guide safer and more rational driving behavior. The versatility of the proposed model, including its potential to incorporate additional derivative constraints and broader architectural applications, enhances its appeal and broadens its impact within the scientific community.
♻ ☆ Uncertainty Quantification for Motor Imagery BCI -- Machine Learning vs. Deep Learning
Brain-computer interfaces (BCIs) turn brain signals into functionally useful output, but they are not always accurate. A good Machine Learning classifier should be able to indicate how confident it is about a given classification, by giving a probability for its classification. Standard classifiers for Motor Imagery BCIs do give such probabilities, but research on uncertainty quantification has been limited to Deep Learning. We compare the uncertainty quantification ability of established BCI classifiers using Common Spatial Patterns (CSP-LDA) and Riemannian Geometry (MDRM) to specialized methods in Deep Learning (Deep Ensembles and Direct Uncertainty Quantification) as well as standard Convolutional Neural Networks (CNNs). We found that the overconfidence typically seen in Deep Learning is not a problem in CSP-LDA and MDRM. We found that MDRM is underconfident, which we solved by adding Temperature Scaling (MDRM-T). CSP-LDA and MDRM-T give the best uncertainty estimates, but Deep Ensembles and standard CNNs give the best classifications. We show that all models are able to separate between easy and difficult estimates, so that we can increase the accuracy of a Motor Imagery BCI by rejecting samples that are ambiguous.
comment: 6 pages, 3 figures; fixed typos
♻ ☆ What's Pulling the Strings? Evaluating Integrity and Attribution in AI Training and Inference through Concept Shift CCS
The growing adoption of artificial intelligence (AI) has amplified concerns about trustworthiness, including integrity, privacy, robustness, and bias. To assess and attribute these threats, we propose ConceptLens, a generic framework that leverages pre-trained multimodal models to identify the root causes of integrity threats by analyzing Concept Shift in probing samples. ConceptLens demonstrates strong detection performance for vanilla data poisoning attacks and uncovers vulnerabilities to bias injection, such as the generation of covert advertisements through malicious concept shifts. It identifies privacy risks in unaltered but high-risk samples, filters them before training, and provides insights into model weaknesses arising from incomplete or imbalanced training data. Additionally, at the model level, it attributes concepts that the target model is overly dependent on, identifies misleading concepts, and explains how disrupting key concepts negatively impacts the model. Furthermore, it uncovers sociological biases in generative content, revealing disparities across sociological contexts. Strikingly, ConceptLens reveals how safe training and inference data can be unintentionally and easily exploited, potentially undermining safety alignment. Our study informs actionable insights to breed trust in AI systems, thereby speeding adoption and driving greater innovation.
comment: Accepted to The ACM Conference on Computer and Communications Security (CCS) 2025
♻ ☆ Structured and Balanced Multi-Component and Multi-Layer Neural Networks
In this work, we propose a balanced multi-component and multi-layer neural network (MMNN) structure to accurately and efficiently approximate functions with complex features, in terms of both degrees of freedom and computational cost. The main idea is inspired by a multi-component approach, in which each component can be effectively approximated by a single-layer network, combined with a multi-layer decomposition strategy to capture the complexity of the target function. Although MMNNs can be viewed as a simple modification of fully connected neural networks (FCNNs) or multi-layer perceptrons (MLPs) by introducing balanced multi-component structures, they achieve a significant reduction in training parameters, a much more efficient training process, and improved accuracy compared to FCNNs or MLPs. Extensive numerical experiments demonstrate the effectiveness of MMNNs in approximating highly oscillatory functions and their ability to automatically adapt to localized features.
comment: Our codes and implementation details are available at https://github.com/ShijunZhangMath/MMNN
♻ ☆ Proactive Agents for Multi-Turn Text-to-Image Generation Under Uncertainty
User prompts for generative AI models are often underspecified, leading to a misalignment between the user intent and models' understanding. As a result, users commonly have to painstakingly refine their prompts. We study this alignment problem in text-to-image (T2I) generation and propose a prototype for proactive T2I agents equipped with an interface to (1) actively ask clarification questions when uncertain, and (2) present their uncertainty about user intent as an understandable and editable belief graph. We build simple prototypes for such agents and propose a new scalable and automated evaluation approach using two agents, one with a ground truth intent (an image) while the other tries to ask as few questions as possible to align with the ground truth. We experiment over three image-text datasets: ImageInWords (Garg et al., 2024), COCO (Lin et al., 2014) and DesignBench, a benchmark we curated with strong artistic and design elements. Experiments over the three datasets demonstrate the proposed T2I agents' ability to ask informative questions and elicit crucial information to achieve successful alignment with at least 2 times higher VQAScore (Lin et al., 2024) than the standard T2I generation. Moreover, we conducted human studies and observed that at least 90% of human subjects found these agents and their belief graphs helpful for their T2I workflow, highlighting the effectiveness of our approach. Code and DesignBench can be found at https://github.com/google-deepmind/proactive_t2i_agents.
♻ ☆ A Thorough Assessment of the Non-IID Data Impact in Federated Learning
Federated learning (FL) allows collaborative machine learning (ML) model training among decentralized clients' information, ensuring data privacy. The decentralized nature of FL deals with non-independent and identically distributed (non-IID) data. This open problem has notable consequences, such as decreased model performance and more significant convergence times. Despite its importance, experimental studies systematically addressing all types of data heterogeneity (a.k.a. non-IIDness) remain scarce. We aim to fill this gap by assessing and quantifying the non-IID effect through a thorough empirical analysis. We use the Hellinger Distance (HD) to measure differences in distribution among clients. Our study benchmarks four state-of-the-art strategies for handling non-IID data, including label, feature, quantity, and spatiotemporal skewness, under realistic and controlled conditions. This is the first comprehensive analysis of the spatiotemporal skew effect in FL. Our findings highlight the significant impact of label and spatiotemporal skew non-IID types on FL model performance, with notable performance drops occurring at specific HD thresholds. Additionally, the FL performance is heavily affected mainly when the non-IIDness is extreme. Thus, we provide recommendations for FL research to tackle data heterogeneity effectively. Our work represents the most extensive examination of non-IIDness in FL, offering a robust foundation for future research.
♻ ☆ FADE: Why Bad Descriptions Happen to Good Features
Recent advances in mechanistic interpretability have highlighted the potential of automating interpretability pipelines in analyzing the latent representations within LLMs. While this may enhance our understanding of internal mechanisms, the field lacks standardized evaluation methods for assessing the validity of discovered features. We attempt to bridge this gap by introducing FADE: Feature Alignment to Description Evaluation, a scalable model-agnostic framework for automatically evaluating feature-to-description alignment. FADE evaluates alignment across four key metrics - Clarity, Responsiveness, Purity, and Faithfulness - and systematically quantifies the causes of the misalignment between features and their descriptions. We apply FADE to analyze existing open-source feature descriptions and assess key components of automated interpretability pipelines, aiming to enhance the quality of descriptions. Our findings highlight fundamental challenges in generating feature descriptions, particularly for SAEs compared to MLP neurons, providing insights into the limitations and future directions of automated interpretability. We release FADE as an open-source package at: https://github.com/brunibrun/FADE
♻ ☆ Holistic analysis on the sustainability of Federated Learning across AI product lifecycle ECAI 2025
In light of emerging legal requirements and policies focused on privacy protection, there is a growing trend of companies across various industries adopting Federated Learning (FL). This decentralized approach involves multiple clients or silos, collaboratively training a global model under the coordination of a central server while utilizing their private local data. Unlike traditional methods that necessitate data sharing and transmission, Cross-Silo FL allows clients to share model updates rather than raw data, thereby enhancing privacy. Despite its growing adoption, the carbon impact associated with Cross-Silo FL remains poorly understood due to the limited research in this area. This study seeks to bridge this gap by evaluating the sustainability of Cross-Silo FL throughout the entire AI product lifecycle, extending the analysis beyond the model training phase alone. We systematically compare this decentralized method with traditional centralized approaches and present a robust quantitative framework for assessing the costs and CO2 emissions in real-world Cross-Silo FL environments. Our findings indicate that the energy consumption and costs of model training are comparable between Cross-Silo Federated Learning and Centralized Learning. However, the additional data transfer and storage requirements inherent in Centralized Learning can result in significant, often overlooked CO2 emissions. Moreover, we introduce an innovative data and application management system that integrates Cross-Silo FL and analytics, aiming at improving the sustainability and economic efficiency of IT enterprises.
comment: Accepted in ECAI 2025 (PAIS track)
♻ ☆ Error bounds for particle gradient descent, and extensions of the log-Sobolev and Talagrand inequalities
We prove non-asymptotic error bounds for particle gradient descent (PGD, Kuntz et al., 2023), a recently introduced algorithm for maximum likelihood estimation of large latent variable models obtained by discretizing a gradient flow of the free energy. We begin by showing that the flow converges exponentially fast to the free energy's minimizers for models satisfying a condition that generalizes both the log-Sobolev and the Polyak--{\L}ojasiewicz inequalities (LSI and P{\L}I, respectively). We achieve this by extending a result well-known in the optimal transport literature (that the LSI implies the Talagrand inequality) and its counterpart in the optimization literature (that the P{\L}I implies the so-called quadratic growth condition), and applying the extension to our new setting. We also generalize the Bakry--\'Emery Theorem and show that the LSI/P{\L}I extension holds for models with strongly concave log-likelihoods. For such models, we further control PGD's discretization error and obtain the non-asymptotic error bounds. While we are motivated by the study of PGD, we believe that the inequalities and results we extend may be of independent interest.
♻ ☆ Prominent Roles of Conditionally Invariant Components in Domain Adaptation: Theory and Algorithms
Domain adaptation (DA) is a statistical learning problem that arises when the distribution of the source data used to train a model differs from that of the target data used to evaluate the model. While many DA algorithms have demonstrated considerable empirical success, blindly applying these algorithms can often lead to worse performance on new datasets. To address this, it is crucial to clarify the assumptions under which a DA algorithm has good target performance. In this work, we focus on the assumption of the presence of conditionally invariant components (CICs), which are relevant for prediction and remain conditionally invariant across the source and target data. We demonstrate that CICs, which can be estimated through conditional invariant penalty (CIP), play three prominent roles in providing target risk guarantees in DA. First, we propose a new algorithm based on CICs, importance-weighted conditional invariant penalty (IW-CIP), which has target risk guarantees beyond simple settings such as covariate shift and label shift. Second, we show that CICs help identify large discrepancies between source and target risks of other DA algorithms. Finally, we demonstrate that incorporating CICs into the domain invariant projection (DIP) algorithm can address its failure scenario caused by label-flipping features. We support our new algorithms and theoretical findings via numerical experiments on synthetic data, MNIST, CelebA, Camelyon17, and DomainNet datasets.
♻ ☆ MTF-Grasp: A Multi-tier Federated Learning Approach for Robotic Grasping
Federated Learning (FL) is a promising machine learning paradigm that enables participating devices to train privacy-preserved and collaborative models. FL has proven its benefits for robotic manipulation tasks. However, grasping tasks lack exploration in such settings where robots train a global model without moving data and ensuring data privacy. The main challenge is that each robot learns from data that is nonindependent and identically distributed (non-IID) and of low quantity. This exhibits performance degradation, particularly in robotic grasping. Thus, in this work, we propose MTF-Grasp, a multi-tier FL approach for robotic grasping, acknowledging the unique challenges posed by the non-IID data distribution across robots, including quantitative skewness. MTF-Grasp harnesses data quality and quantity across robots to select a set of "top-level" robots with better data distribution and higher sample count. It then utilizes top-level robots to train initial seed models and distribute them to the remaining "low-level" robots, reducing the risk of model performance degradation in low-level robots. Our approach outperforms the conventional FL setup by up to 8% on the quantity-skewed Cornell and Jacquard grasping datasets.
comment: The work is accepted for presentation at IEEE SMC 2025
♻ ☆ 2.5D Object Detection for Intelligent Roadside Infrastructure SC
On-board sensors of autonomous vehicles can be obstructed, occluded, or limited by restricted fields of view, complicating downstream driving decisions. Intelligent roadside infrastructure perception systems, installed at elevated vantage points, can provide wide, unobstructed intersection coverage, supplying a complementary information stream to autonomous vehicles via vehicle-to-everything (V2X) communication. However, conventional 3D object-detection algorithms struggle to generalize under the domain shift introduced by top-down perspectives and steep camera angles. We introduce a 2.5D object detection framework, tailored specifically for infrastructure roadside-mounted cameras. Unlike conventional 2D or 3D object detection, we employ a prediction approach to detect ground planes of vehicles as parallelograms in the image frame. The parallelogram preserves the planar position, size, and orientation of objects while omitting their height, which is unnecessary for most downstream applications. For training, a mix of real-world and synthetically generated scenes is leveraged. We evaluate generalizability on a held-out camera viewpoint and in adverse-weather scenarios absent from the training set. Our results show high detection accuracy, strong cross-viewpoint generalization, and robustness to diverse lighting and weather conditions. Model weights and inference code are provided at: https://gitlab.kit.edu/kit/aifb/ATKS/public/digit4taf/2.5d-object-detection
comment: Accepted at 2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC)
♻ ☆ LHU-Net: a Lean Hybrid U-Net for Cost-efficient, High-performance Volumetric Segmentation MICCAI-2025
The rise of Transformer architectures has advanced medical image segmentation, leading to hybrid models that combine Convolutional Neural Networks (CNNs) and Transformers. However, these models often suffer from excessive complexity and fail to effectively integrate spatial and channel features, crucial for precise segmentation. To address this, we propose LHU-Net, a Lean Hybrid U-Net for volumetric medical image segmentation. LHU-Net prioritizes spatial feature extraction before refining channel features, optimizing both efficiency and accuracy. Evaluated on four benchmark datasets (Synapse, Left Atrial, BraTS-Decathlon, and Lung-Decathlon), LHU-Net consistently outperforms existing models across diverse modalities (CT/MRI) and output configurations. It achieves state-of-the-art Dice scores while using four times fewer parameters and 20% fewer FLOPs than competing models, without the need for pre-training, additional data, or model ensembles. With an average of 11 million parameters, LHU-Net sets a new benchmark for computational efficiency and segmentation accuracy. Our implementation is available on GitHub: https://github.com/xmindflow/LHUNet
comment: Accepted at MICCAI-2025
♻ ☆ NeuTSFlow: Modeling Continuous Functions Behind Time Series Forecasting
Time series forecasting is a fundamental task with broad applications, yet conventional methods often treat data as discrete sequences, overlooking their origin as noisy samples of continuous processes. Crucially, discrete noisy observations cannot uniquely determine a continuous function; instead, they correspond to a family of plausible functions. Mathematically, time series can be viewed as noisy observations of a continuous function family governed by a shared probability measure. Thus, the forecasting task can be framed as learning the transition from the historical function family to the future function family. This reframing introduces two key challenges: (1) How can we leverage discrete historical and future observations to learn the relationships between their underlying continuous functions? (2) How can we model the transition path in function space from the historical function family to the future function family? To address these challenges, we propose NeuTSFlow, a novel framework that leverages Neural Operators to facilitate flow matching for learning path of measure between historical and future function families. By parameterizing the velocity field of the flow in infinite-dimensional function spaces, NeuTSFlow moves beyond traditional methods that focus on dependencies at discrete points, directly modeling function-level features instead. Experiments on diverse forecasting tasks demonstrate NeuTSFlow's superior accuracy and robustness, validating the effectiveness of the function-family perspective.
♻ ☆ Sharing is CAIRing: Characterizing Principles and Assessing Properties of Universal Privacy Evaluation for Synthetic Tabular Data
Data sharing is a necessity for innovative progress in many domains, especially in healthcare. However, the ability to share data is hindered by regulations protecting the privacy of natural persons. Synthetic tabular data provide a promising solution to address data sharing difficulties but does not inherently guarantee privacy. Still, there is a lack of agreement on appropriate methods for assessing the privacy-preserving capabilities of synthetic data, making it difficult to compare results across studies. To the best of our knowledge, this is the first work to identify properties that constitute good universal privacy evaluation metrics for synthetic tabular data. The goal of universally applicable metrics is to enable comparability across studies and to allow non-technical stakeholders to understand how privacy is protected. We identify four principles for the assessment of metrics: Comparability, Applicability, Interpretability, and Representativeness (CAIR). To quantify and rank the degree to which evaluation metrics conform to the CAIR principles, we design a rubric using a scale of 1-4. Each of the four properties is scored on four parameters, yielding 16 total dimensions. We study the applicability and usefulness of the CAIR principles and rubric by assessing a selection of metrics popular in other studies. The results provide granular insights into the strengths and weaknesses of existing metrics that not only rank the metrics but highlight areas of potential improvements. We expect that the CAIR principles will foster agreement among researchers and organizations on which universal privacy evaluation metrics are appropriate for synthetic tabular data.
♻ ☆ Governance of Generative Artificial Intelligence for Companies
Generative Artificial Intelligence (GenAI), specifically large language models(LLMs) like ChatGPT, has swiftly entered organizations without adequate governance, posing both opportunities and risks. Despite extensive debates on GenAI's transformative nature and regulatory measures, limited research addresses organizational governance, encompassing technical and business perspectives. Although numerous frameworks for governance of AI exist, it is not clear to what extent they apply to GenAI. Our review paper fills this gap by surveying recent works with the purpose of better understanding fundamental characteristics of GenAI and adjusting prior frameworks specifically towards GenAI governance within companies. To do so, it extends Nickerson's framework development processes to include prior conceptualizations. Our framework outlines the scope, objectives, and governance mechanisms tailored to harness business opportunities as well as mitigate risks associated with GenAI integration. Our research contributes a focused approach to GenAI governance, offering practical insights for companies navigating the challenges of GenAI adoption and highlighting research gaps.
comment: This paper is under submission
♻ ☆ Patherea: Cell Detection and Classification for the 2020s
We present Patherea, a unified framework for point-based cell detection and classification that enables the development and fair evaluation of state-of-the-art methods. To support this, we introduce a large-scale dataset that replicates the clinical workflow for Ki-67 proliferation index estimation. Our method directly predicts cell locations and classes without relying on intermediate representations. It incorporates a hybrid Hungarian matching strategy for accurate point assignment and supports flexible backbones and training regimes, including recent pathology foundation models. Patherea achieves state-of-the-art performance on public datasets - Lizard, BRCA-M2C, and BCData - while highlighting performance saturation on these benchmarks. In contrast, our newly proposed Patherea dataset presents a significantly more challenging benchmark. Additionally, we identify and correct common errors in current evaluation protocols and provide an updated benchmarking utility for standardized assessment. The Patherea dataset and code are publicly available to facilitate further research and fair comparisons.
comment: Submitted to Medical Image Analysis
♻ ☆ Protecting Copyrighted Material with Unique Identifiers in Large Language Model Training
A primary concern regarding training large language models (LLMs) is whether they abuse copyrighted online text. With the increasing training data scale and the prevalence of LLMs in daily lives, two problems arise: \textbf{1)} false positive membership inference results misled by similar examples; \textbf{2)} membership inference methods are usually too complex for end users to understand and use. To address these issues, we propose an alternative \textit{insert-and-detect} methodology, advocating that web users and content platforms employ \textbf{\textit{unique identifiers}} for reliable and independent membership inference. Users and platforms can create their identifiers, embed them in copyrighted text, and independently detect them in future LLMs. As an initial demonstration, we introduce \textit{\textbf{ghost sentences}} and a user-friendly last-$k$ words test, allowing end users to chat with LLMs for membership inference. Ghost sentences consist primarily of unique passphrases of random natural words, which can come with customized elements to bypass possible filter rules. The last-$k$ words test requires a significant repetition time of ghost sentences~($\ge10$). For cases with fewer repetitions, we designed an extra perplexity test, as LLMs exhibit high perplexity when encountering unnatural passphrases. We also conduct a comprehensive study on the memorization and membership inference of ghost sentences, examining factors such as training data scales, model sizes, repetition times, insertion positions, wordlist of passphrases, alignment, \textit{etc}. Our study shows the possibility of applying ghost sentences in real scenarios and provides instructions for the potential application.
comment: A technical report, work mainly done in the early of 2024
♻ ☆ Data Augmentation in Time Series Forecasting through Inverted Framework
Currently, iTransformer is one of the most popular and effective models for multivariate time series (MTS) forecasting. Thanks to its inverted framework, iTransformer effectively captures multivariate correlation. However, the inverted framework still has some limitations. It diminishes temporal interdependency information, and introduces noise in cases of nonsignificant variable correlation. To address these limitations, we introduce a novel data augmentation method on inverted framework, called DAIF. Unlike previous data augmentation methods, DAIF stands out as the first real-time augmentation specifically designed for the inverted framework in MTS forecasting. We first define the structure of the inverted sequence-to-sequence framework, then propose two different DAIF strategies, Frequency Filtering and Cross-variation Patching to address the existing challenges of the inverted framework. Experiments across multiple datasets and inverted models have demonstrated the effectiveness of our DAIF.
comment: The paper is under consideration at Pattern Recognition Letters
♻ ☆ Complexity-Aware Training of Deep Neural Networks for Optimal Structure Discovery
We propose a novel algorithm for combined unit and layer pruning of deep neural networks that functions during training and without requiring a pre-trained network to apply. Our algorithm optimally trades-off learning accuracy and pruning levels while balancing layer vs. unit pruning and computational vs. parameter complexity using only three user-defined parameters, which are easy to interpret and tune. We formulate a stochastic optimization problem over the network weights and the parameters of variational Bernoulli distributions for binary Random Variables taking values either 0 or 1 and scaling the units and layers of the network. Optimal network structures are found as the solution to this optimization problem. Pruning occurs when a variational parameter converges to 0 rendering the corresponding structure permanently inactive, thus saving computations both during training and prediction. A key contribution of our approach is to define a cost function that combines the objectives of prediction accuracy and network pruning in a computational/parameter complexity-aware manner and the automatic selection of the many regularization parameters. We show that the proposed algorithm converges to solutions of the optimization problem corresponding to deterministic networks. We analyze the ODE system that underlies our stochastic optimization algorithm and establish domains of attraction for the dynamics of the network parameters. These theoretical results lead to practical pruning conditions avoiding the premature pruning of units and layers during training. We evaluate our method on the CIFAR-10/100 and ImageNet datasets using ResNet architectures and demonstrate that it gives improved results with respect to pruning ratios and test accuracy over layer-only or unit-only pruning and favorably competes with combined unit and layer pruning algorithms requiring pre-trained networks.
comment: 31 pages, 6 figures, 6 tables. Restructured Sections 1 and 3, added simulation data in Section 5 and added appendices
♻ ☆ FedRef: Communication-Efficient Bayesian Fine Tuning with Reference Model
Federated learning(FL) is used for distributed scenarios to train artificial intelligence(AI) models while ensuring users' privacy. In federated learning scenario, the server generally never knows about users' data. This type of concept makes the AI training process efficient in terms of data privacy. However, regarding model performance, federated AI models may not sufficiently satisfy AI users' expectations. Furthermore, AI users have a wide range of different needs. It is not easy to satisfy the whole users needs. These types of issues can be addressed through AI model optimization, fine-tuning, or personalization to achieve optimal model performance. To address model optimization challenges, we propose reference model-based federated learning for optimal fine-tuning, which overcomes catastrophic forgetting in each round. This method is derived from Bayesian parameter-efficient transfer learning, which includes an optimal proximal term and utilizes a reference model that incorporates previous model parameters. As a result, this method achieves both high model performance and clients' low computing cost.
comment: 6 pages,14 equation, 4 figure, 1table
♻ ☆ From Observational Data to Clinical Recommendations: A Causal Framework for Estimating Patient-level Treatment Effects and Learning Policies
We propose a framework for building patient-specific treatment recommendation models, building on the large recent literature on learning patient-level causal models and inspired by the target trial paradigm of Hernan and Robins. We focus on safety and validity, including the crucial issue of causal identification when using observational data. We do not provide a specific model, but rather a way to integrate existing methods and know-how into a practical pipeline. We further provide a real world use-case of treatment optimization for patients with heart failure who develop acute kidney injury during hospitalization. The results suggest our pipeline can improve patient outcomes over the current treatment regime.
♻ ☆ Learning to Reason at the Frontier of Learnability
Reinforcement learning is now widely adopted as the final stage of large language model training, especially for reasoning-style tasks such as maths problems. Typically, models attempt each question many times during a single training step and attempt to learn from their successes and failures. However, we demonstrate that throughout training with two popular algorithms (PPO and VinePPO) on two widely used datasets, many questions are either solved by all attempts - meaning they are already learned - or by none - providing no meaningful training signal. To address this, we adapt a method from the reinforcement learning literature - sampling for learnability - and apply it to the reinforcement learning stage of LLM training. Our curriculum prioritises questions with high variance of success, i.e. those where the agent sometimes succeeds, but not always. Our findings demonstrate that this curriculum consistently boosts training performance across multiple algorithms and datasets, paving the way for more efficient and effective reinforcement learning with LLMs.
♻ ☆ The late-stage training dynamics of (stochastic) subgradient descent on homogeneous neural networks COLT 2025
We analyze the implicit bias of constant step stochastic subgradient descent (SGD). We consider the setting of binary classification with homogeneous neural networks - a large class of deep neural networks with ReLU-type activation functions such as MLPs and CNNs without biases. We interpret the dynamics of normalized SGD iterates as an Euler-like discretization of a conservative field flow that is naturally associated to the normalized classification margin. Owing to this interpretation, we show that normalized SGD iterates converge to the set of critical points of the normalized margin at late-stage training (i.e., assuming that the data is correctly classified with positive normalized margin). Up to our knowledge, this is the first extension of the analysis of Lyu and Li (2020) on the discrete dynamics of gradient descent to the nonsmooth and stochastic setting. Our main result applies to binary classification with exponential or logistic losses. We additionally discuss extensions to more general settings.
comment: Accepted/presented at the 38th Annual Conference on Learning Theory (COLT 2025)
♻ ☆ Enhancing RLHF with Human Gaze Modeling
Reinforcement Learning from Human Feedback (RLHF) aligns language models with human preferences but is computationally expensive. We explore two approaches that leverage human gaze modeling to enhance RLHF: (1) gaze-aware reward models and (2) gaze-based distribution of sparse rewards at token level. Our experiments demonstate that gaze-informed RLHF achieves faster convergence while maintaining or slightly improving performance, thus, reducing computational costs during policy optimization. These results show that human gaze provides a valuable and underused signal for policy optimization, pointing to a promising direction for improving RLHF efficiency.
♻ ☆ A Computational Theory and Semi-Supervised Algorithm for Clustering
A computational theory for clustering and a semi-supervised clustering algorithm is presented. Clustering is defined to be the obtainment of groupings of data such that each group contains no anomalies with respect to a chosen grouping principle and measure; all other examples are considered to be fringe points, isolated anomalies, anomalous clusters or unknown clusters. More precisely, after appropriate modelling under the assumption of uniform random distribution, any example whose expectation of occurrence is <1 with respect to a group is considered an anomaly; otherwise it is assigned a membership of that group. Thus, clustering is conceived as the dual of anomaly detection. The representation of data is taken to be the Euclidean distance of a point to a cluster median. This is due to the robustness properties of the median to outliers, its approximate location of centrality and so that decision boundaries are general purpose. The kernel of the clustering method is the perception anomaly detection algorithm, resulting in a parameter-free, fast, and efficient clustering algorithm. Acknowledging that clustering is an interactive and iterative process, the algorithm relies on a small fraction of known relationships between examples. These relationships serve as seeds to define the user's objectives and guide the clustering process. The method then expands the clusters accordingly, leaving the remaining examples for exploration and subsequent iterations. Results are presented on synthetic and realworld data sets, demonstrating the advantages over the most popular unsupervised and semi-supervised clustering methods.
♻ ☆ How does Watermarking Affect Visual Language Models in Document Understanding?
Visual Language Models (VLMs) have become foundational models for document understanding tasks, widely used in the processing of complex multimodal documents across domains such as finance, law, and academia. However, documents often contain noise-like information, such as watermarks, which inevitably leads us to inquire: \emph{Do watermarks degrade the performance of VLMs in document understanding?} To address this, we propose a novel evaluation framework to investigate the effect of visible watermarks on VLMs performance. We takes into account various factors, including different types of document data, the positions of watermarks within documents and variations in watermark content. Our experimental results reveal that VLMs performance can be significantly compromised by watermarks, with performance drop rates reaching up to 36\%. We discover that \emph{scattered} watermarks cause stronger interference than centralized ones, and that \emph{semantic contents} in watermarks creates greater disruption than simple visual occlusion. Through attention mechanism analysis and embedding similarity examination, we find that the performance drops are mainly attributed to that watermarks 1) force widespread attention redistribution, and 2) alter semantic representation in the embedding space. Our research not only highlights significant challenges in deploying VLMs for document understanding, but also provides insights towards developing robust inference mechanisms on watermarked documents.
comment: Accepted to COLM 2025
♻ ☆ Labels Generated by Large Language Models Help Measure People's Empathy in Vitro
Large language models (LLMs) have revolutionised many fields, with LLM-as-a-service (LLMSaaS) offering accessible, general-purpose solutions without costly task-specific training. In contrast to the widely studied prompt engineering for directly solving tasks (in vivo), this paper explores LLMs' potential for in-vitro applications: using LLM-generated labels to improve supervised training of mainstream models. We examine two strategies - (1) noisy label correction and (2) training data augmentation - in empathy computing, an emerging task to predict psychology-based questionnaire outcomes from inputs like textual narratives. Crowdsourced datasets in this domain often suffer from noisy labels that misrepresent underlying empathy. We show that replacing or supplementing these crowdsourced labels with LLM-generated labels, developed using psychology-based scale-aware prompts, achieves statistically significant accuracy improvements. Notably, the RoBERTa pre-trained language model (PLM) trained with noise-reduced labels yields a state-of-the-art Pearson correlation coefficient of 0.648 on the public NewsEmp benchmarks. This paper further analyses evaluation metric selection and demographic biases to help guide the future development of more equitable empathy computing models. Code and LLM-generated labels are available at https://github.com/hasan-rakibul/LLMPathy.
comment: This work has been submitted to the IEEE for possible publication
♻ ☆ Predictable Scale: Part I, Step Law -- Optimal Hyperparameter Scaling Law in Large Language Model Pretraining
The impressive capabilities of Large Language Models (LLMs) across diverse tasks are now well\text{-}established, yet their effective deployment necessitates careful hyperparameter optimization. Although existing methods have explored the influence of hyperparameters on model performance, a principled and generalizable framework across model architectures and data recipes remains absent. In this study, we conduct an unprecedented empirical investigation\text{-} training over 3,700 LLMs from scratch across 100 trillion tokens, consuming nearly one million NVIDIA H800 GPU hours to establish a universal Scaling Law for hyperparameter optimization in LLM Pre-training, called \textbf{Step Law}. We empirically observe that, under fixed model size ($N$) and dataset size ($D$), the hyperparameter landscape exhibits convexity with a broad optimum, substantially reducing the complexity of hyperparameter search. Building on this insight, we formally define and empirically validate the Step Law: The optimal learning rate follows a power-law relationship with $N$ and $D$, while the optimal batch size is primarily influenced by $D$ and remains largely invariant to $N$.Notably, our estimated optima deviate from the global best performance found via exhaustive search by merely \textbf{0.094\%} on the test set. To our best known, Step Law is the \textbf{first} that unifies different model shapes and structures, such as Mixture-of-Experts models and dense transformers, as well as establishes optimal hyperparameter scaling laws across diverse data recipes. We contribute a universal, plug-and-play optimal hyperparameter tool for the community, which is expected to advance efficient LLM training at scale. All experimental code, data and checkpoints are publicly available at \href{https://github.com/step-law/steplaw}{https://github.com/step-law/steplaw}.
comment: 22 pages
♻ ☆ Predictable Scale: Part II, Farseer: A Refined Scaling Law in Large Language Models
Training Large Language Models (LLMs) is prohibitively expensive, creating a critical scaling gap where insights from small-scale experiments often fail to transfer to resource-intensive production systems, thereby hindering efficient innovation. To bridge this, we introduce Farseer, a novel and refined scaling law offering enhanced predictive accuracy across scales. By systematically constructing a model loss surface $L(N,D)$, Farseer achieves a significantly better fit to empirical data than prior laws (e.g., Chinchilla's law). Our methodology yields accurate, robust, and highly generalizable predictions, demonstrating excellent extrapolation capabilities, improving upon Chinchilla's law by reducing extrapolation error by 433\%. This allows for the reliable evaluation of competing training strategies across all $(N,D)$ settings, enabling conclusions from small-scale ablation studies to be confidently extrapolated to predict large-scale performance. Furthermore, Farseer provides new insights into optimal compute allocation, better reflecting the nuanced demands of modern LLM training. To validate our approach, we trained an extensive suite of approximately 1,000 LLMs across diverse scales and configurations, consuming roughly 3 million NVIDIA H100 GPU hours. We are comprehensively open-sourcing all models, data, results, and logs at https://github.com/Farseer-Scaling-Law/Farseer to foster further research.
comment: 34
♻ ☆ Regret Analysis of Posterior Sampling-Based Expected Improvement for Bayesian Optimization
Bayesian optimization is a powerful tool for optimizing an expensive-to-evaluate black-box function. In particular, the effectiveness of expected improvement (EI) has been demonstrated in a wide range of applications. However, theoretical analyses of EI are limited compared with other theoretically established algorithms. This paper analyzes a randomized variant of EI, which evaluates the EI from the maximum of the posterior sample path. We show that this posterior sampling-based random EI achieves the sublinear Bayesian cumulative regret bounds under the assumption that the black-box function follows a Gaussian process. Finally, we demonstrate the effectiveness of the proposed method through numerical experiments.
comment: 35pages, 5 figures, fix trivial errors
♻ ☆ Simplifying Graph Kernels for Efficient
While kernel methods and Graph Neural Networks offer complementary strengths, integrating the two has posed challenges in efficiency and scalability. The Graph Neural Tangent Kernel provides a theoretical bridge by interpreting GNNs through the lens of neural tangent kernels. However, its reliance on deep, stacked layers introduces repeated computations that hinder performance. In this work, we introduce a new perspective by designing the simplified graph kernel, which replaces deep layer stacking with a streamlined $K$-step message aggregation process. This formulation avoids iterative layer-wise propagation altogether, leading to a more concise and computationally efficient framework without sacrificing the expressive power needed for graph tasks. Beyond this simplification, we propose another Simplified Graph Kernel, which draws from Gaussian Process theory to model infinite-width GNNs. Rather than simulating network depth, this kernel analytically computes kernel values based on the statistical behavior of nonlinear activations in the infinite limit. This eliminates the need for explicit architecture simulation, further reducing complexity. Our experiments on standard graph and node classification benchmarks show that our methods achieve competitive accuracy while reducing runtime. This makes them practical alternatives for learning on graphs at scale. Full implementation and reproducibility materials are provided at: https://anonymous.4open.science/r/SGNK-1CE4/.
♻ ☆ Decoder-Hybrid-Decoder Architecture for Efficient Reasoning with Long Generation
Recent advances in language modeling have demonstrated the effectiveness of State Space Models (SSMs) for efficient sequence modeling. While hybrid architectures such as Samba and the decoder-decoder architecture, YOCO, have shown promising performance gains over Transformers, prior works have not investigated the efficiency potential of representation sharing between SSM layers. In this paper, we introduce the Gated Memory Unit (GMU), a simple yet effective mechanism for efficient memory sharing across layers. We apply it to create SambaY, a decoder-hybrid-decoder architecture that incorporates GMUs in the cross-decoder to share memory readout states from a Samba-based self-decoder. SambaY significantly enhances decoding efficiency, preserves linear pre-filling time complexity, and boosts long-context performance, all while eliminating the need for explicit positional encoding. Through extensive scaling experiments, we demonstrate that our model exhibits a significantly lower irreducible loss compared to a strong YOCO baseline, indicating superior performance scalability under large-scale compute regimes. Our largest model enhanced with Differential Attention, Phi4-mini-Flash-Reasoning, achieves significantly better performance than Phi4-mini-Reasoning on reasoning tasks such as Math500, AIME24/25, and GPQA Diamond without any reinforcement learning, while delivering up to 10x higher decoding throughput on 2K-length prompts with 32K generation length under the vLLM inference framework. We release our training codebase on open-source data at https://github.com/microsoft/ArchScale.
♻ ☆ PATCH: a deep learning method to assess heterogeneity of artistic practice in historical paintings
The history of art has seen significant shifts in the manner in which artworks are created, making understanding of creative processes a central question in technical art history. In the Renaissance and Early Modern period, paintings were largely produced by master painters directing workshops of apprentices who often contributed to projects. The masters varied significantly in artistic and managerial styles, meaning different combinations of artists and implements might be seen both between masters and within workshops or even individual canvases. Information on how different workshops were managed and the processes by which artworks were created remains elusive. Machine learning methods have potential to unearth new information about artists' creative processes by extending the analysis of brushwork to a microscopic scale. Analysis of workshop paintings, however, presents a challenge in that documentation of the artists and materials involved is sparse, meaning external examples are not available to train networks to recognize their contributions. Here we present a novel machine learning approach we call pairwise assignment training for classifying heterogeneity (PATCH) that is capable of identifying individual artistic practice regimes with no external training data, or "ground truth." The method achieves unsupervised results by supervised means, and outperforms both simple statistical procedures and unsupervised machine learning methods. We apply this method to two historical paintings by the Spanish Renaissance master, El Greco: The Baptism of Christ and Christ on the Cross with Landscape, and our findings regarding the former potentially challenge previous work that has assigned the painting to workshop members. Further, the results of our analyses create a measure of heterogeneity of artistic practice that can be used to characterize artworks across time and space.
comment: main text: 15 pages, 5 figures; SI: 10 pages, 4 figures; v2: minor typo corrections, higher resolution figures; v3: additional comparisons with alternative methods
♻ ☆ BRIDGE: Bootstrapping Text to Control Time-Series Generation via Multi-Agent Iterative Optimization and Diffusion Modeling ICML 2025
Time-series Generation (TSG) is a prominent research area with broad applications in simulations, data augmentation, and counterfactual analysis. While existing methods have shown promise in unconditional single-domain TSG, real-world applications demand for cross-domain approaches capable of controlled generation tailored to domain-specific constraints and instance-level requirements. In this paper, we argue that text can provide semantic insights, domain information and instance-specific temporal patterns, to guide and improve TSG. We introduce ``Text-Controlled TSG'', a task focused on generating realistic time series by incorporating textual descriptions. To address data scarcity in this setting, we propose a novel LLM-based Multi-Agent framework that synthesizes diverse, realistic text-to-TS datasets. Furthermore, we introduce BRIDGE, a hybrid text-controlled TSG framework that integrates semantic prototypes with text description for supporting domain-level guidance. This approach achieves state-of-the-art generation fidelity on 11 of 12 datasets, and improves controllability by up to 12% on MSE and 6% MAE compared to no text input generation, highlighting its potential for generating tailored time-series data.
comment: ICML 2025 Main Conference
♻ ☆ Rethinking Data Protection in the (Generative) Artificial Intelligence Era
The (generative) artificial intelligence (AI) era has profoundly reshaped the meaning and value of data. No longer confined to static content, data now permeates every stage of the AI lifecycle from the training samples that shape model parameters to the prompts and outputs that drive real-world model deployment. This shift renders traditional notions of data protection insufficient, while the boundaries of what needs safeguarding remain poorly defined. Failing to safeguard data in AI systems can inflict societal and individual, underscoring the urgent need to clearly delineate the scope of and rigorously enforce data protection. In this perspective, we propose a four-level taxonomy, including non-usability, privacy preservation, traceability, and deletability, that captures the diverse protection needs arising in modern (generative) AI models and systems. Our framework offers a structured understanding of the trade-offs between data utility and control, spanning the entire AI pipeline, including training datasets, model weights, system prompts, and AI-generated content. We analyze representative technical approaches at each level and reveal regulatory blind spots that leave critical assets exposed. By offering a structured lens to align future AI technologies and governance with trustworthy data practices, we underscore the urgency of rethinking data protection for modern AI techniques and provide timely guidance for developers, researchers, and regulators alike.
comment: Perspective paper for a broader scientific audience. The first two authors contributed equally to this paper. 13 pages
♻ ☆ Tuning Algorithmic and Architectural Hyperparameters in Graph-Based Semi-Supervised Learning with Provable Guarantees UAI 2025
Graph-based semi-supervised learning is a powerful paradigm in machine learning for modeling and exploiting the underlying graph structure that captures the relationship between labeled and unlabeled data. A large number of classical as well as modern deep learning based algorithms have been proposed for this problem, often having tunable hyperparameters. We initiate a formal study of tuning algorithm hyperparameters from parameterized algorithm families for this problem. We obtain novel $O(\log n)$ pseudo-dimension upper bounds for hyperparameter selection in three classical label propagation-based algorithm families, where $n$ is the number of nodes, implying bounds on the amount of data needed for learning provably good parameters. We further provide matching $\Omega(\log n)$ pseudo-dimension lower bounds, thus asymptotically characterizing the learning-theoretic complexity of the parameter tuning problem. We extend our study to selecting architectural hyperparameters in modern graph neural networks. We bound the Rademacher complexity for tuning the self-loop weighting in recently proposed Simplified Graph Convolution (SGC) networks. We further propose a tunable architecture that interpolates graph convolutional neural networks (GCN) and graph attention networks (GAT) in every layer, and provide Rademacher complexity bounds for tuning the interpolation coefficient.
comment: 31 pages (12 pages main body), 2 figures. UAI 2025
♻ ☆ Truncated Kernel Stochastic Gradient Descent on Spheres
Inspired by the structure of spherical harmonics, we propose the truncated kernel stochastic gradient descent (T-kernel SGD) algorithm with a least-square loss function for spherical data fitting. T-kernel SGD introduces a novel regularization strategy by implementing stochastic gradient descent through a closed-form solution of the projection of the stochastic gradient in a low-dimensional subspace. In contrast to traditional kernel SGD, the regularization strategy implemented by T-kernel SGD is more effective in balancing bias and variance by dynamically adjusting the hypothesis space during iterations. The most significant advantage of the proposed algorithm is that it can achieve theoretically optimal convergence rates using a constant step size (independent of the sample size) while overcoming the inherent saturation problem of kernel SGD. Additionally, we leverage the structure of spherical polynomials to derive an equivalent T-kernel SGD, significantly reducing storage and computational costs compared to kernel SGD. Typically, T-kernel SGD requires only $\mathcal{O}(n^{1+\frac{d}{d-1}\epsilon})$ computational complexity and $\mathcal{O}(n^{\frac{d}{d-1}\epsilon})$ storage to achieve optimal rates for the d-dimensional sphere, where $0<\epsilon<\frac{1}{2}$ can be arbitrarily small if the optimal fitting or the underlying space possesses sufficient regularity. This regularity is determined by the smoothness parameter of the objective function and the decaying rate of the eigenvalues of the integral operator associated with the kernel function, both of which reflect the difficulty of the estimation problem. Our main results quantitatively characterize how this prior information influences the convergence of T-kernel SGD. The numerical experiments further validate the theoretical findings presented in this paper.
comment: 74 pages, 22 figures
♻ ☆ TextDestroyer: A Training- and Annotation-Free Diffusion Method for Destroying Anomal Text from Images
In this paper, we propose TextDestroyer, the first training- and annotation-free method for scene text destruction using a pre-trained diffusion model. Existing scene text removal models require complex annotation and retraining, and may leave faint yet recognizable text information, compromising privacy protection and content concealment. TextDestroyer addresses these issues by employing a three-stage hierarchical process to obtain accurate text masks. Our method scrambles text areas in the latent start code using a Gaussian distribution before reconstruction. During the diffusion denoising process, self-attention key and value are referenced from the original latent to restore the compromised background. Latent codes saved at each inversion step are used for replacement during reconstruction, ensuring perfect background restoration. The advantages of TextDestroyer include: (1) it eliminates labor-intensive data annotation and resource-intensive training; (2) it achieves more thorough text destruction, preventing recognizable traces; and (3) it demonstrates better generalization capabilities, performing well on both real-world scenes and generated images.
♻ ☆ Can Prompt Difficulty be Online Predicted for Accelerating RL Finetuning of Reasoning Models?
Recent advances have witnessed the effectiveness of reinforcement learning (RL) finetuning in enhancing the reasoning capabilities of large language models (LLMs). The optimization process often requires numerous iterations to achieve satisfactory performance, resulting in high computational costs due to the need for frequent prompt evaluations under intensive LLM interactions and repeated policy updates. Appropriate online prompt selection methods reduce iteration steps by prioritizing informative prompts during training, while the pipeline's reliance on exhaustive prompt evaluation and subset selection for optimization still incurs substantial computational overhead due to frequent LLM inference calls. Distinguished from these direct evaluate-then-select schemes, this work investigates iterative approximate evaluation for arbitrary prompts and introduces Model Predictive Prompt Selection (MoPPS), a Bayesian risk-predictive framework that online estimates prompt difficulty without requiring costly LLM interactions. Technically, MoPPS models each prompt's success rate as a latent variable, performs streaming Bayesian inference, and employs posterior sampling in a constructed multi-armed bandit machine, enabling sample efficient and adaptive prompt selection. Extensive experiments across mathematics, planning, and vision-based geometry tasks show that MoPPS reliably predicts prompt difficulty and accelerates training with significantly reduced LLM rollouts.
♻ ☆ Learning Time-Varying Multi-Region Brain Communications via Scalable Markovian Gaussian Processes
Understanding and constructing brain communications that capture dynamic communications across multiple regions is fundamental to modern system neuroscience, yet current methods struggle to find time-varying region-level communications or scale to large neural datasets with long recording durations. We present a novel framework using Markovian Gaussian Processes to learn brain communications with time-varying temporal delays from multi-region neural recordings, named Adaptive Delay Model (ADM). Our method combines Gaussian Processes with State Space Models and employs parallel scan inference algorithms, enabling efficient scaling to large datasets while identifying concurrent communication patterns that evolve over time. This time-varying approach captures how brain region interactions shift dynamically during cognitive processes. Validated on synthetic and multi-region neural recordings datasets, our approach discovers both the directionality and temporal dynamics of neural communication. This work advances our understanding of distributed neural computation and provides a scalable tool for analyzing dynamic brain networks.
♻ ☆ Epic-Sounds: A Large-scale Dataset of Actions That Sound
We introduce EPIC-SOUNDS, a large-scale dataset of audio annotations capturing temporal extents and class labels within the audio stream of the egocentric videos. We propose an annotation pipeline where annotators temporally label distinguishable audio segments and describe the action that could have caused this sound. We identify actions that can be discriminated purely from audio, through grouping these free-form descriptions of audio into classes. For actions that involve objects colliding, we collect human annotations of the materials of these objects (e.g. a glass object being placed on a wooden surface), which we verify from video, discarding ambiguities. Overall, EPIC-SOUNDS includes 78.4k categorised segments of audible events and actions, distributed across 44 classes as well as 39.2k non-categorised segments. We train and evaluate state-of-the-art audio recognition and detection models on our dataset, for both audio-only and audio-visual methods. We also conduct analysis on: the temporal overlap between audio events, the temporal and label correlations between audio and visual modalities, the ambiguities in annotating materials from audio-only input, the importance of audio-only labels and the limitations of current models to understand actions that sound.
comment: Accepted at TPAMI
♻ ☆ HueManity: Probing Fine-Grained Visual Perception in MLLMs
Multimodal Large Language Models (MLLMs) excel at high-level visual reasoning, but their performance on nuanced perceptual tasks remains surprisingly limited. We present HueManity, a benchmark designed to assess visual perception in MLLMs. The dataset comprises 83,850 images featuring two-character alphanumeric strings embedded in Ishihara test style dot patterns, challenging models on precise pattern recognition. Our evaluation of nine state-of-the-art MLLMs on HueManity demonstrates a significant performance deficit compared to human and traditional computer vision baselines. The best-performing MLLM achieved a 33.6% accuracy on the numeric `easy' task and a striking 3% on the alphanumeric `hard' task. In contrast, human participants achieved near-perfect scores (100% and 95.6%), and a fine-tuned ResNet50 model reached accuracies of 96.5% and 94.5%. These results highlight a critical gap in the visual capabilities of current MLLMs. Our analysis further explores potential architectural and training-paradigm factors contributing to this perceptual gap in MLLMs. We open-source HueManity dataset and code to foster further research in improving perceptual robustness of MLLMs.
♻ ☆ GeoChain: Multimodal Chain-of-Thought for Geographic Reasoning
This paper introduces GeoChain, a large-scale benchmark for evaluating step-by-step geographic reasoning in multimodal large language models (MLLMs). Leveraging 1.46 million Mapillary street-level images, GeoChain pairs each image with a 21-step chain-of-thought (CoT) question sequence (over 30 million Q&A pairs). These sequences guide models from coarse attributes to fine-grained localization across four reasoning categories - visual, spatial, cultural, and precise geolocation - annotated by difficulty. Images are also enriched with semantic segmentation (150 classes) and a visual locatability score. Our benchmarking of contemporary MLLMs (GPT-4.1 variants, Claude 3.7, Gemini 2.5 variants) on a diverse 2,088-image subset reveals consistent challenges: models frequently exhibit weaknesses in visual grounding, display erratic reasoning, and struggle to achieve accurate localization, especially as the reasoning complexity escalates. GeoChain offers a robust diagnostic methodology, critical for fostering significant advancements in complex geographic reasoning within MLLMs.
♻ ☆ METIS: Fast Quality-Aware RAG Systems with Configuration Adaptation
RAG (Retrieval Augmented Generation) allows LLMs (large language models) to generate better responses with external knowledge, but using more external knowledge often improves generation quality at the expense of response delay. Prior work either reduces the response delay (through better scheduling of RAG queries) or strives to maximize quality (which involves tuning the RAG workflow), but they fall short in optimizing the tradeoff between the delay and quality of RAG responses. This paper presents METIS, the first RAG system that jointly schedules queries and adapts the key RAG configurations of each query, such as the number of retrieved text chunks and synthesis methods, in order to balance quality optimization and response delay reduction. Using 4 popular RAG-QA datasets, we show that compared with the state-of-the-art RAG optimization schemes, METIS reduces the generation latency by $1.64-2.54\times$ without sacrificing generation quality.
comment: 17 pages, 18 figures
Multimedia 7
☆ InstructFLIP: Exploring Unified Vision-Language Model for Face Anti-spoofing
Face anti-spoofing (FAS) aims to construct a robust system that can withstand diverse attacks. While recent efforts have concentrated mainly on cross-domain generalization, two significant challenges persist: limited semantic understanding of attack types and training redundancy across domains. We address the first by integrating vision-language models (VLMs) to enhance the perception of visual input. For the second challenge, we employ a meta-domain strategy to learn a unified model that generalizes well across multiple domains. Our proposed InstructFLIP is a novel instruction-tuned framework that leverages VLMs to enhance generalization via textual guidance trained solely on a single domain. At its core, InstructFLIP explicitly decouples instructions into content and style components, where content-based instructions focus on the essential semantics of spoofing, and style-based instructions consider variations related to the environment and camera characteristics. Extensive experiments demonstrate the effectiveness of InstructFLIP by outperforming SOTA models in accuracy and substantially reducing training redundancy across diverse domains in FAS. Project website is available at https://kunkunlin1221.github.io/InstructFLIP.
comment: Accepted by MM'25
☆ Stereo Sound Event Localization and Detection with Onscreen/offscreen Classification
This paper presents the objective, dataset, baseline, and metrics of Task 3 of the DCASE2025 Challenge on sound event localization and detection (SELD). In previous editions, the challenge used four-channel audio formats of first-order Ambisonics (FOA) and microphone array. In contrast, this year's challenge investigates SELD with stereo audio data (termed stereo SELD). This change shifts the focus from more specialized 360{\deg} audio and audiovisual scene analysis to more commonplace audio and media scenarios with limited field-of-view (FOV). Due to inherent angular ambiguities in stereo audio data, the task focuses on direction-of-arrival (DOA) estimation in the azimuth plane (left-right axis) along with distance estimation. The challenge remains divided into two tracks: audio-only and audiovisual, with the audiovisual track introducing a new sub-task of onscreen/offscreen event classification necessitated by the limited FOV. This challenge introduces the DCASE2025 Task3 Stereo SELD Dataset, whose stereo audio and perspective video clips are sampled and converted from the STARSS23 recordings. The baseline system is designed to process stereo audio and corresponding video frames as inputs. In addition to the typical SELD event classification and localization, it integrates onscreen/offscreen classification for the audiovisual track. The evaluation metrics have been modified to introduce an onscreen/offscreen accuracy metric, which assesses the models' ability to identify which sound sources are onscreen. In the experimental evaluation, the baseline system performs reasonably well with the stereo audio data.
comment: 5 pages, 2 figures
☆ POLYCHARTQA: Benchmarking Large Vision-Language Models with Multilingual Chart Question Answering
Charts are a universally adopted medium for interpreting and communicating data. However, existing chart understanding benchmarks are predominantly English-centric, limiting their accessibility and applicability to global audiences. In this paper, we present PolyChartQA, the first large-scale multilingual chart question answering benchmark covering 22,606 charts and 26,151 question-answering pairs across 10 diverse languages. PolyChartQA is built using a decoupled pipeline that separates chart data from rendering code, allowing multilingual charts to be flexibly generated by simply translating the data and reusing the code. We leverage state-of-the-art LLM-based translation and enforce rigorous quality control in the pipeline to ensure the linguistic and semantic consistency of the generated multilingual charts. PolyChartQA facilitates systematic evaluation of multilingual chart understanding. Experiments on both open- and closed-source large vision-language models reveal a significant performance gap between English and other languages, especially low-resource ones with non-Latin scripts. This benchmark lays a foundation for advancing globally inclusive vision-language models.
comment: Work in Progress
☆ Unveiling the Visual Rhetoric of Persuasive Cartography: A Case Study of the Design of Octopus Maps
When designed deliberately, data visualizations can become powerful persuasive tools, influencing viewers' opinions, values, and actions. While researchers have begun studying this issue (e.g., to evaluate the effects of persuasive visualization), we argue that a fundamental mechanism of persuasion resides in rhetorical construction, a perspective inadequately addressed in current visualization research. To fill this gap, we present a focused analysis of octopus maps, a visual genre that has maintained persuasive power across centuries and achieved significant social impact. Employing rhetorical schema theory, we collected and analyzed 90 octopus maps spanning from the 19th century to contemporary times. We closely examined how octopus maps implement their persuasive intents and constructed a design space that reveals how visual metaphors are strategically constructed and what common rhetorical strategies are applied to components such as maps, octopus imagery, and text. Through the above analysis, we also uncover a set of interesting findings. For instance, contrary to the common perception that octopus maps are primarily a historical phenomenon, our research shows that they remain a lively design convention in today's digital age. Additionally, while most octopus maps stem from Western discourse that views the octopus as an evil symbol, some designs offer alternative interpretations, highlighting the dynamic nature of rhetoric across different sociocultural settings. Lastly, drawing from the lessons provided by octopus maps, we discuss the associated ethical concerns of persuasive visualization.
☆ Catching Dark Signals in Algorithms: Unveiling Audiovisual and Thematic Markers of Unsafe Content Recommended for Children and Teenagers
The prevalence of short form video platforms, combined with the ineffectiveness of age verification mechanisms, raises concerns about the potential harms facing children and teenagers in an algorithm-moderated online environment. We conducted multimodal feature analysis and thematic topic modeling of 4,492 short videos recommended to children and teenagers on Instagram Reels, TikTok, and YouTube Shorts, collected as a part of an algorithm auditing experiment. This feature-level and content-level analysis revealed that unsafe (i.e., problematic, mentally distressing) short videos (a) possess darker visual features and (b) contain explicitly harmful content and implicit harm from anxiety-inducing ordinary content. We introduce a useful framework of online harm (i.e., explicit, implicit, unintended), providing a unique lens for understanding the dynamic, multifaceted online risks facing children and teenagers. The findings highlight the importance of protecting younger audiences in critical developmental stages from both explicit and implicit risks on social media, calling for nuanced content moderation, age verification, and platform regulation.
☆ Wavelet-GS: 3D Gaussian Splatting with Wavelet Decomposition
3D Gaussian Splatting (3DGS) has revolutionized 3D scene reconstruction, which effectively balances rendering quality, efficiency, and speed. However, existing 3DGS approaches usually generate plausible outputs and face significant challenges in complex scene reconstruction, manifesting as incomplete holistic structural outlines and unclear local lighting effects. To address these issues simultaneously, we propose a novel decoupled optimization framework, which integrates wavelet decomposition into 3D Gaussian Splatting and 2D sampling. Technically, through 3D wavelet decomposition, our approach divides point clouds into high-frequency and low-frequency components, enabling targeted optimization for each. The low-frequency component captures global structural outlines and manages the distribution of Gaussians through voxelization. In contrast, the high-frequency component restores intricate geometric and textural details while incorporating a relight module to mitigate lighting artifacts and enhance photorealistic rendering. Additionally, a 2D wavelet decomposition is applied to the training images, simulating radiance variations. This provides critical guidance for high-frequency detail reconstruction, ensuring seamless integration of details with the global structure. Extensive experiments on challenging datasets demonstrate our method achieves state-of-the-art performance across various metrics, surpassing existing approaches and advancing the field of 3D scene reconstruction.
comment: 9 pages
♻ ☆ Token Communications: A Large Model-Driven Framework for Cross-modal Context-aware Semantic Communications
In this paper, we introduce token communications (TokCom), a large model-driven framework to leverage cross-modal context information in generative semantic communications (GenSC). TokCom is a new paradigm, motivated by the recent success of generative foundation models and multimodal large language models (GFM/MLLMs), where the communication units are tokens, enabling efficient transformer-based token processing at the transmitter and receiver. In this paper, we introduce the potential opportunities and challenges of leveraging context in GenSC, explore how to integrate GFM/MLLMs-based token processing into semantic communication systems to leverage cross-modal context effectively at affordable complexity, present the key principles for efficient TokCom at various layers in future wireless networks. In a typical image semantic communication setup, we demonstrate a significant improvement of the bandwidth efficiency, achieved by TokCom by leveraging the context information among tokens. Finally, the potential research directions are identified to facilitate adoption of TokCom in future wireless networks.
comment: Accepted at IEEE Wireless Communications Magazine
Computation and Language 108
☆ AirLLM: Diffusion Policy-based Adaptive LoRA for Remote Fine-Tuning of LLM over the Air
Operating Large Language Models (LLMs) on edge devices is increasingly challenged by limited communication bandwidth and strained computational and memory costs. Thus, cloud-assisted remote fine-tuning becomes indispensable. Nevertheless, existing Low-Rank Adaptation (LoRA) approaches typically employ fixed or heuristic rank configurations, and the subsequent over-the-air transmission of all LoRA parameters could be rather inefficient. To address this limitation, we develop AirLLM, a hierarchical diffusion policy framework for communication-aware LoRA adaptation. Specifically, AirLLM models the rank configuration as a structured action vector that spans all LoRA-inserted projections. To solve the underlying high-dimensional sequential decision-making problem, a Proximal Policy Optimization (PPO) agent generates coarse-grained decisions by jointly observing wireless states and linguistic complexity, which are then refined via Denoising Diffusion Implicit Models (DDIM) to produce high-resolution, task- and channel-adaptive rank vectors. The two modules are optimized alternatively, with the DDIM trained under the Classifier-Free Guidance (CFG) paradigm to maintain alignment with PPO rewards. Experiments under varying signal-to-noise ratios demonstrate that AirLLM consistently enhances fine-tuning performance while significantly reducing transmission costs, highlighting the effectiveness of reinforcement-driven, diffusion-refined rank adaptation for scalable and efficient remote fine-tuning over the air.
comment: 11 pages, 8 figures
☆ Real-World Summarization: When Evaluation Reaches Its Limits
We examine evaluation of faithfulness to input data in the context of hotel highlights: brief LLM-generated summaries that capture unique features of accommodations. Through human evaluation campaigns involving categorical error assessment and span-level annotation, we compare traditional metrics, trainable methods, and LLM-as-a-judge approaches. Our findings reveal that simpler metrics like word overlap correlate surprisingly well with human judgments (Spearman correlation rank of 0.63), often outperforming more complex methods when applied to out-of-domain data. We further demonstrate that while LLMs can generate high-quality highlights, they prove unreliable for evaluation as they tend to severely under- or over-annotate. Our analysis of real-world business impacts shows incorrect and non-checkable information pose the greatest risks. We also highlight challenges in crowdsourced evaluations.
☆ Reasoning Strategies in Large Language Models: Can They Follow, Prefer, and Optimize?
Human reasoning involves different strategies, each suited to specific problems. Prior work shows that large language model (LLMs) tend to favor a single reasoning strategy, potentially limiting their effectiveness in diverse reasoning challenges. In this work, we investigate whether prompting can control LLMs reasoning strategies and assess its impact on logical problem-solving. While our experiments show that no single strategy consistently improves accuracy, performance could be enhanced if models could adaptively choose the optimal strategy. We propose methods to guide LLMs in strategy selection, highlighting new ways to refine their reasoning abilities.
☆ Seq vs Seq: An Open Suite of Paired Encoders and Decoders
The large language model (LLM) community focuses almost exclusively on decoder-only language models, since they are easier to use for text generation. However, a large subset of the community still uses encoder-only models for tasks such as classification or retrieval. Previous work has attempted to compare these architectures, but is forced to make comparisons with models that have different numbers of parameters, training techniques, and datasets. We introduce the SOTA open-data Ettin suite of models: paired encoder-only and decoder-only models ranging from 17 million parameters to 1 billion, trained on up to 2 trillion tokens. Using the same recipe for both encoder-only and decoder-only models produces SOTA recipes in both categories for their respective sizes, beating ModernBERT as an encoder and Llama 3.2 and SmolLM2 as decoders. Like previous work, we find that encoder-only models excel at classification and retrieval tasks while decoders excel at generative tasks. However, we show that adapting a decoder model to encoder tasks (and vice versa) through continued training is subpar compared to using only the reverse objective (i.e. a 400M encoder outperforms a 1B decoder on MNLI, and vice versa for generative tasks). We open-source all artifacts of this study including training data, training order segmented by checkpoint, and 200+ checkpoints to allow future work to analyze or extend all aspects of training.
☆ KisMATH: Do LLMs Have Knowledge of Implicit Structures in Mathematical Reasoning?
Chain-of-thought traces have been shown to improve performance of large language models in a plethora of reasoning tasks, yet there is no consensus on the mechanism through which this performance boost is achieved. To shed more light on this, we introduce Causal CoT Graphs (CCGs), which are directed acyclic graphs automatically extracted from reasoning traces that model fine-grained causal dependencies in the language model output. A collection of $1671$ mathematical reasoning problems from MATH500, GSM8K and AIME, and their associated CCGs are compiled into our dataset -- \textbf{KisMATH}. Our detailed empirical analysis with 15 open-weight LLMs shows that (i) reasoning nodes in the CCG are mediators for the final answer, a condition necessary for reasoning; and (ii) LLMs emphasise reasoning paths given by the CCG, indicating that models internally realise structures akin to our graphs. KisMATH enables controlled, graph-aligned interventions and opens up avenues for further investigation into the role of chain-of-thought in LLM reasoning.
comment: 15 pages, 9 figures
☆ EXAONE 4.0: Unified Large Language Models Integrating Non-reasoning and Reasoning Modes
This technical report introduces EXAONE 4.0, which integrates a Non-reasoning mode and a Reasoning mode to achieve both the excellent usability of EXAONE 3.5 and the advanced reasoning abilities of EXAONE Deep. To pave the way for the agentic AI era, EXAONE 4.0 incorporates essential features such as agentic tool use, and its multilingual capabilities are extended to support Spanish in addition to English and Korean. The EXAONE 4.0 model series consists of two sizes: a mid-size 32B model optimized for high performance, and a small-size 1.2B model designed for on-device applications. The EXAONE 4.0 demonstrates superior performance compared to open-weight models in its class and remains competitive even against frontier-class models. The models are publicly available for research purposes and can be easily downloaded via https://huggingface.co/LGAI-EXAONE.
comment: Technical Report, 30 Pages
☆ DCR: Quantifying Data Contamination in LLMs Evaluation
The rapid advancement of large language models (LLMs) has heightened concerns about benchmark data contamination (BDC), where models inadvertently memorize evaluation data, inflating performance metrics and undermining genuine generalization assessment. This paper introduces the Data Contamination Risk (DCR) framework, a lightweight, interpretable pipeline designed to detect and quantify BDC across four granular levels: semantic, informational, data, and label. By synthesizing contamination scores via a fuzzy inference system, DCR produces a unified DCR Factor that adjusts raw accuracy to reflect contamination-aware performance. Validated on 9 LLMs (0.5B-72B) across sentiment analysis, fake news detection, and arithmetic reasoning tasks, the DCR framework reliably diagnoses contamination severity and with accuracy adjusted using the DCR Factor to within 4% average error across the three benchmarks compared to the uncontaminated baseline. Emphasizing computational efficiency and transparency, DCR provides a practical tool for integrating contamination assessment into routine evaluations, fostering fairer comparisons and enhancing the credibility of LLM benchmarking practices.
☆ Addressing Data Imbalance in Transformer-Based Multi-Label Emotion Detection with Weighted Loss SemEval 2025
This paper explores the application of a simple weighted loss function to Transformer-based models for multi-label emotion detection in SemEval-2025 Shared Task 11. Our approach addresses data imbalance by dynamically adjusting class weights, thereby enhancing performance on minority emotion classes without the computational burden of traditional resampling methods. We evaluate BERT, RoBERTa, and BART on the BRIGHTER dataset, using evaluation metrics such as Micro F1, Macro F1, ROC-AUC, Accuracy, and Jaccard similarity coefficients. The results demonstrate that the weighted loss function improves performance on high-frequency emotion classes but shows limited impact on minority classes. These findings underscore both the effectiveness and the challenges of applying this approach to imbalanced multi-label emotion detection.
comment: 10 pages, 1 figure, SemEval 2025
☆ What is the Best Process Model Representation? A Comparative Analysis for Process Modeling with Large Language Models
Large Language Models (LLMs) are increasingly applied for Process Modeling (PMo) tasks such as Process Model Generation (PMG). To support these tasks, researchers have introduced a variety of Process Model Representations (PMRs) that serve as model abstractions or generation targets. However, these PMRs differ widely in structure, complexity, and usability, and have never been systematically compared. Moreover, recent PMG approaches rely on distinct evaluation strategies and generation techniques, making comparison difficult. This paper presents the first empirical study that evaluates multiple PMRs in the context of PMo with LLMs. We introduce the PMo Dataset, a new dataset containing 55 process descriptions paired with models in nine different PMRs. We evaluate PMRs along two dimensions: suitability for LLM-based PMo and performance on PMG. \textit{Mermaid} achieves the highest overall score across six PMo criteria, whereas \textit{BPMN text} delivers the best PMG results in terms of process element similarity.
comment: 12 pages, 7 figures, to be published in AI4BPM 2025 Proceedings
☆ Automated Novelty Evaluation of Academic Paper: A Collaborative Approach Integrating Human and Large Language Model Knowledge
Novelty is a crucial criterion in the peer review process for evaluating academic papers. Traditionally, it's judged by experts or measure by unique reference combinations. Both methods have limitations: experts have limited knowledge, and the effectiveness of the combination method is uncertain. Moreover, it's unclear if unique citations truly measure novelty. The large language model (LLM) possesses a wealth of knowledge, while human experts possess judgment abilities that the LLM does not possess. Therefore, our research integrates the knowledge and abilities of LLM and human experts to address the limitations of novelty assessment. The most common novelty in academic papers is the introduction of new methods. In this paper, we propose leveraging human knowledge and LLM to assist pretrained language models (PLMs, e.g. BERT etc.) in predicting the method novelty of papers. Specifically, we extract sentences related to the novelty of the academic paper from peer review reports and use LLM to summarize the methodology section of the academic paper, which are then used to fine-tune PLMs. In addition, we have designed a text-guided fusion module with novel Sparse-Attention to better integrate human and LLM knowledge. We compared the method we proposed with a large number of baselines. Extensive experiments demonstrate that our method achieves superior performance.
comment: Journal of the Association for Information Science and Technology, 2025
☆ Internal Value Alignment in Large Language Models through Controlled Value Vector Activation ACL 2025
Aligning Large Language Models (LLMs) with human values has attracted increasing attention since it provides clarity, transparency, and the ability to adapt to evolving scenarios. In this paper, we introduce a Controlled Value Vector Activation (ConVA) method that directly aligns the internal values of LLMs by interpreting how a value is encoded in their latent representations and modifies relevant activations to ensure consistent values in LLMs. To ensure an accurate and unbiased interpretation, we propose a context-controlled value vector identification method. To consistently control values without sacrificing model performance, we introduce a gated value vector activation method for effective and minimum degree of value control. Experiments show that our method achieves the highest control success rate across 10 basic values without hurting LLM performance and fluency, and ensures target values even with opposite and potentially malicious input prompts. Source code and data are available at~ https://github.com/hr-jin/ConVA.
comment: 25 pages, 14 figures. Accepted by ACL 2025 (main conference)
☆ Dr.Copilot: A Multi-Agent Prompt Optimized Assistant for Improving Patient-Doctor Communication in Romanian
Text-based telemedicine has become increasingly common, yet the quality of medical advice in doctor-patient interactions is often judged more on how advice is communicated rather than its clinical accuracy. To address this, we introduce Dr.Copilot , a multi-agent large language model (LLM) system that supports Romanian-speaking doctors by evaluating and enhancing the presentation quality of their written responses. Rather than assessing medical correctness, Dr.Copilot provides feedback along 17 interpretable axes. The system comprises of three LLM agents with prompts automatically optimized via DSPy. Designed with low-resource Romanian data and deployed using open-weight models, it delivers real-time specific feedback to doctors within a telemedicine platform. Empirical evaluations and live deployment with 41 doctors show measurable improvements in user reviews and response quality, marking one of the first real-world deployments of LLMs in Romanian medical settings.
comment: 10 figures, 2 tables, 2 listings
☆ Fine-Grained Chinese Hate Speech Understanding: Span-Level Resources, Coded Term Lexicon, and Enhanced Detection Frameworks
The proliferation of hate speech has inflicted significant societal harm, with its intensity and directionality closely tied to specific targets and arguments. In recent years, numerous machine learning-based methods have been developed to detect hateful comments on online platforms automatically. However, research on Chinese hate speech detection lags behind, and interpretability studies face two major challenges: first, the scarcity of span-level fine-grained annotated datasets limits models' deep semantic understanding of hate speech; second, insufficient research on identifying and interpreting coded hate speech restricts model explainability in complex real-world scenarios. To address these, we make the following contributions: (1) We introduce the Span-level Target-Aware Toxicity Extraction dataset (STATE ToxiCN), the first span-level Chinese hate speech dataset, and evaluate the hate semantic understanding of existing models using it. (2) We conduct the first comprehensive study on Chinese coded hate terms, LLMs' ability to interpret hate semantics. (3) We propose a method to integrate an annotated lexicon into models, significantly enhancing hate speech detection performance. Our work provides valuable resources and insights to advance the interpretability of Chinese hate speech detection research.
☆ FMC: Formalization of Natural Language Mathematical Competition Problems ICML 2025
Efficient and accurate autoformalization methods, which leverage large-scale datasets of extensive natural language mathematical problems to construct formal language datasets, are key to advancing formal mathematical reasoning. In this paper, we propose an autoformalization pipeline based on large language models with error feedback, achieving a fully automatic and training-free formalization approach. Using this pipeline, we curate an Olympiad-level dataset aligning natural language problems with Lean formalizations. The dataset comprises $3,922$ mathematical problems in natural language and $9,787$ in Lean, of which $64.46\%$ were assessed as at least above-average quality, making it suitable as a benchmark for automated theorem provers. Additionally, we investigate the formalization and reasoning capabilities of various LLMs and empirically demonstrate that few-shot learning, error feedback, and increasing sampling numbers enhance the autoformalization process. Experiments of three automated theorem provers on the \dataset\ dataset also highlight its challenging nature and its value as a benchmark for formal reasoning tasks.
comment: Accepted in ICML 2025 AI4MATH Workshop
☆ KV-Latent: Dimensional-level KV Cache Reduction with Frequency-aware Rotary Positional Embedding ACL 2025
Large language models (LLMs) based on Transformer Decoders have become the preferred choice for conversational generative AI. Despite the overall superiority of the Decoder architecture, the gradually increasing Key-Value (KV) cache during inference has emerged as a primary efficiency bottleneck, both in aspects of memory consumption and data transfer bandwidth limitations. To address these challenges, we propose a paradigm called KV-Latent. By down-sampling the Key-Value vector dimensions into a latent space, we can significantly reduce the KV Cache footprint and improve inference speed, only with a small amount of extra training, less than 1\% of pre-training takes. Besides, we enhanced the stability of Rotary Positional Embedding applied on lower-dimensional vectors by modifying its frequency sampling mechanism, avoiding noise introduced by higher frequencies while retaining position attenuation. Our experiments, including both models with Grouped Query Attention and those without, have yielded satisfactory results. Finally, we conducted comparative experiments to study the impact of separately reducing Key and Value components on model's performance. Our approach allows for the construction of more efficient language model systems, and opens the new possibility on KV Cache saving and efficient LLMs. Our code is available at https://github.com/ShiLuohe/KV-Latent.
comment: To be published in The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)
☆ Sparse Autoencoders Can Capture Language-Specific Concepts Across Diverse Languages
Understanding the multilingual mechanisms of large language models (LLMs) provides insight into how they process different languages, yet this remains challenging. Existing studies often focus on individual neurons, but their polysemantic nature makes it difficult to isolate language-specific units from cross-lingual representations. To address this, we explore sparse autoencoders (SAEs) for their ability to learn monosemantic features that represent concrete and abstract concepts across languages in LLMs. While some of these features are language-independent, the presence of language-specific features remains underexplored. In this work, we introduce SAE-LAPE, a method based on feature activation probability, to identify language-specific features within the feed-forward network. We find that many such features predominantly appear in the middle to final layers of the model and are interpretable. These features influence the model's multilingual performance and language output and can be used for language identification with performance comparable to fastText along with more interpretability. Our code is available at https://github.com/LyzanderAndrylie/language-specific-features .
☆ An Agentic Flow for Finite State Machine Extraction using Prompt Chaining
Finite-State Machines (FSMs) are critical for modeling the operational logic of network protocols, enabling verification, analysis, and vulnerability discovery. However, existing FSM extraction techniques face limitations such as scalability, incomplete coverage, and ambiguity in natural language specifications. In this paper, we propose FlowFSM, a novel agentic framework that leverages Large Language Models (LLMs) combined with prompt chaining and chain-of-thought reasoning to extract accurate FSMs from raw RFC documents. FlowFSM systematically processes protocol specifications, identifies state transitions, and constructs structured rule-books by chaining agent outputs. Experimental evaluation across FTP and RTSP protocols demonstrates that FlowFSM achieves high extraction precision while minimizing hallucinated transitions, showing promising results. Our findings highlight the potential of agent-based LLM systems in the advancement of protocol analysis and FSM inference for cybersecurity and reverse engineering applications.
☆ EsBBQ and CaBBQ: The Spanish and Catalan Bias Benchmarks for Question Answering
Previous literature has largely shown that Large Language Models (LLMs) perpetuate social biases learnt from their pre-training data. Given the notable lack of resources for social bias evaluation in languages other than English, and for social contexts outside of the United States, this paper introduces the Spanish and the Catalan Bias Benchmarks for Question Answering (EsBBQ and CaBBQ). Based on the original BBQ, these two parallel datasets are designed to assess social bias across 10 categories using a multiple-choice QA setting, now adapted to the Spanish and Catalan languages and to the social context of Spain. We report evaluation results on different LLMs, factoring in model family, size and variant. Our results show that models tend to fail to choose the correct answer in ambiguous scenarios, and that high QA accuracy often correlates with greater reliance on social biases.
☆ Temperature and Persona Shape LLM Agent Consensus With Minimal Accuracy Gains in Qualitative Coding
Large Language Models (LLMs) enable new possibilities for qualitative research at scale, including coding and data annotation. While multi-agent systems (MAS) can emulate human coding workflows, their benefits over single-agent coding remain poorly understood. We conducted an experimental study of how agent persona and temperature shape consensus-building and coding accuracy of dialog segments based on a codebook with 8 codes. Our open-source MAS mirrors deductive human coding through structured agent discussion and consensus arbitration. Using six open-source LLMs (with 3 to 32 billion parameters) and 18 experimental configurations, we analyze over 77,000 coding decisions against a gold-standard dataset of human-annotated transcripts from online math tutoring sessions. Temperature significantly impacted whether and when consensus was reached across all six LLMs. MAS with multiple personas (including neutral, assertive, or empathetic), significantly delayed consensus in four out of six LLMs compared to uniform personas. In three of those LLMs, higher temperatures significantly diminished the effects of multiple personas on consensus. However, neither temperature nor persona pairing lead to robust improvements in coding accuracy. Single agents matched or outperformed MAS consensus in most conditions. Only one model (OpenHermesV2:7B) and code category showed above-chance gains from MAS deliberation when temperature was 0.5 or lower and especially when the agents included at least one assertive persona. Qualitative analysis of MAS collaboration for these configurations suggests that MAS may nonetheless aid in narrowing ambiguous code applications that could improve codebooks and human-AI coding. We contribute new insight into the limits of LLM-based qualitative methods, challenging the notion that diverse MAS personas lead to better outcomes. We open-source our MAS and experimentation code.
comment: Manuscript submitted for review
☆ What Should LLMs Forget? Quantifying Personal Data in LLMs for Right-to-Be-Forgotten Requests KDD 2025
Large Language Models (LLMs) can memorize and reveal personal information, raising concerns regarding compliance with the EU's GDPR, particularly the Right to Be Forgotten (RTBF). Existing machine unlearning methods assume the data to forget is already known but do not address how to identify which individual-fact associations are stored in the model. Privacy auditing techniques typically operate at the population level or target a small set of identifiers, limiting applicability to individual-level data inquiries. We introduce WikiMem, a dataset of over 5,000 natural language canaries covering 243 human-related properties from Wikidata, and a model-agnostic metric to quantify human-fact associations in LLMs. Our approach ranks ground-truth values against counterfactuals using calibrated negative log-likelihood across paraphrased prompts. We evaluate 200 individuals across 15 LLMs (410M-70B parameters), showing that memorization correlates with subject web presence and model scale. We provide a foundation for identifying memorized personal data in LLMs at the individual level, enabling the dynamic construction of forget sets for machine unlearning and RTBF requests.
comment: 16 pages, 3 figures. Accepted at the 7th Workshop on eXplainable Knowledge Discovery in Data Mining (XKDD 2025), ECML PKDD 2025, Porto, Portugal
☆ MSA at ImageCLEF 2025 Multimodal Reasoning: Multilingual Multimodal Reasoning With Ensemble Vision Language Models
We present a robust ensemble-based system for multilingual multimodal reasoning, designed for the ImageCLEF 2025 EXAMS V challenge. Our approach integrates Gemini 2.5 Flash for visual description, Gemini 1.5 Pro for caption refinement and consistency checks, and Gemini 2.5 Pro as a reasoner which handles final answer selection, all coordinated through carefully engineered few-shot and zero-shot prompts. We conducted an extensive ablation study, training several large language models (Gemini 2.5 Flash, Phi 4, Gemma 3, Mistral) on an English dataset and its multilingual augmented version. Additionally, we evaluated Gemini 2.5 Flash in a zero-shot setting for comparison and found it to substantially outperform the trained models. Prompt design also proved critical: enforcing concise, language-normalized formats and prohibiting explanatory text boosted model accuracy on the English validation set from 55.9% to 61.7%. On the official leaderboard, our system (Team MSA) achieved first place overall in the multilingual track with 81.4% accuracy, and led 11 out of 13 individual language tracks, with top results such as 95.07% for Croatian and 92.12% for Italian. These findings highlight that lightweight OCR-VLM ensembles, when paired with precise prompt strategies and cross-lingual augmentation, can outperform heavier end-to-end models in high-stakes, multilingual educational settings.
☆ Multi-Trigger Poisoning Amplifies Backdoor Vulnerabilities in LLMs
Recent studies have shown that Large Language Models (LLMs) are vulnerable to data poisoning attacks, where malicious training examples embed hidden behaviours triggered by specific input patterns. However, most existing works assume a phrase and focus on the attack's effectiveness, offering limited understanding of trigger mechanisms and how multiple triggers interact within the model. In this paper, we present a framework for studying poisoning in LLMs. We show that multiple distinct backdoor triggers can coexist within a single model without interfering with each other, enabling adversaries to embed several triggers concurrently. Using multiple triggers with high embedding similarity, we demonstrate that poisoned triggers can achieve robust activation even when tokens are substituted or separated by long token spans. Our findings expose a broader and more persistent vulnerability surface in LLMs. To mitigate this threat, we propose a post hoc recovery method that selectively retrains specific model components based on a layer-wise weight difference analysis. Our method effectively removes the trigger behaviour with minimal parameter updates, presenting a practical and efficient defence against multi-trigger poisoning.
☆ The Devil behind the mask: An emergent safety vulnerability of Diffusion LLMs
Diffusion-based large language models (dLLMs) have recently emerged as a powerful alternative to autoregressive LLMs, offering faster inference and greater interactivity via parallel decoding and bidirectional modeling. However, despite strong performance in code generation and text infilling, we identify a fundamental safety concern: existing alignment mechanisms fail to safeguard dLLMs against context-aware, masked-input adversarial prompts, exposing novel vulnerabilities. To this end, we present DIJA, the first systematic study and jailbreak attack framework that exploits unique safety weaknesses of dLLMs. Specifically, our proposed DIJA constructs adversarial interleaved mask-text prompts that exploit the text generation mechanisms of dLLMs, i.e., bidirectional modeling and parallel decoding. Bidirectional modeling drives the model to produce contextually consistent outputs for masked spans, even when harmful, while parallel decoding limits model dynamic filtering and rejection sampling of unsafe content. This causes standard alignment mechanisms to fail, enabling harmful completions in alignment-tuned dLLMs, even when harmful behaviors or unsafe instructions are directly exposed in the prompt. Through comprehensive experiments, we demonstrate that DIJA significantly outperforms existing jailbreak methods, exposing a previously overlooked threat surface in dLLM architectures. Notably, our method achieves up to 100% keyword-based ASR on Dream-Instruct, surpassing the strongest prior baseline, ReNeLLM, by up to 78.5% in evaluator-based ASR on JailbreakBench and by 37.7 points in StrongREJECT score, while requiring no rewriting or hiding of harmful content in the jailbreak prompt. Our findings underscore the urgent need for rethinking safety alignment in this emerging class of language models. Code is available at https://github.com/ZichenWen1/DIJA.
comment: 21 pages, 9 figures, work in progress
☆ Beyond Traditional Algorithms: Leveraging LLMs for Accurate Cross-Border Entity Identification
The growing prevalence of cross-border financial activities in global markets has underscored the necessity of accurately identifying and classifying foreign entities. This practice is essential within the Spanish financial system for ensuring robust risk management, regulatory adherence, and the prevention of financial misconduct. This process involves a labor-intensive entity-matching task, where entities need to be validated against available reference sources. Challenges arise from linguistic variations, special characters, outdated names, and changes in legal forms, complicating traditional matching algorithms like Jaccard, cosine, and Levenshtein distances. These methods struggle with contextual nuances and semantic relationships, leading to mismatches. To address these limitations, we explore Large Language Models (LLMs) as a flexible alternative. LLMs leverage extensive training to interpret context, handle abbreviations, and adapt to legal transitions. We evaluate traditional methods, Hugging Face-based LLMs, and interface-based LLMs (e.g., Microsoft Copilot, Alibaba's Qwen 2.5) using a dataset of 65 Portuguese company cases. Results show traditional methods achieve accuracies over 92% but suffer high false positive rates (20-40%). Interface-based LLMs outperform, achieving accuracies above 93%, F1 scores exceeding 96%, and lower false positives (40-80%).
☆ Social Media Sentiments Analysis on the July Revolution in Bangladesh: A Hybrid Transformer Based Machine Learning Approach ECAI 2025
The July Revolution in Bangladesh marked a significant student-led mass uprising, uniting people across the nation to demand justice, accountability, and systemic reform. Social media platforms played a pivotal role in amplifying public sentiment and shaping discourse during this historic mass uprising. In this study, we present a hybrid transformer-based sentiment analysis framework to decode public opinion expressed in social media comments during and after the revolution. We used a brand new dataset of 4,200 Bangla comments collected from social media. The framework employs advanced transformer-based feature extraction techniques, including BanglaBERT, mBERT, XLM-RoBERTa, and the proposed hybrid XMB-BERT, to capture nuanced patterns in textual data. Principle Component Analysis (PCA) were utilized for dimensionality reduction to enhance computational efficiency. We explored eleven traditional and advanced machine learning classifiers for identifying sentiments. The proposed hybrid XMB-BERT with the voting classifier achieved an exceptional accuracy of 83.7% and outperform other model classifier combinations. This study underscores the potential of machine learning techniques to analyze social sentiment in low-resource languages like Bangla.
comment: This paper has been accepted and presented at the IEEE ECAI 2025. The final version will be available in the IEEE Xplore Digital Library
☆ SWE-MERA: A Dynamic Benchmark for Agenticly Evaluating Large Language Models on Software Engineering Tasks
The rapid advancement of Large Language Models (LLMs) in software engineering has revealed critical limitations in existing benchmarks, particularly the widely used SWE-bench dataset. Recent studies have uncovered severe data contamination issues, e.g. SWE-bench reports 32.67% of successful patches involve direct solution leakage and 31.08\% pass due to inadequate test cases. We introduce SWE-MERA, a dynamic, continuously updated benchmark designed to address these fundamental challenges through an automated collection of real-world GitHub issues and rigorous quality validation. Our approach implements a reliable pipeline that ensures quality while minimizing contamination risks, resulting in approximately 10,000 potential tasks with 300 samples currently available. Evaluation using the Aider coding agent demonstrates strong discriminative power in state-of-the-art models. We report performance across a dozen recent LLMs evaluated on tasks collected between September 2024 and June 2025.
☆ LLM-Augmented Symptom Analysis for Cardiovascular Disease Risk Prediction: A Clinical NLP
Timely identification and accurate risk stratification of cardiovascular disease (CVD) remain essential for reducing global mortality. While existing prediction models primarily leverage structured data, unstructured clinical notes contain valuable early indicators. This study introduces a novel LLM-augmented clinical NLP pipeline that employs domain-adapted large language models for symptom extraction, contextual reasoning, and correlation from free-text reports. Our approach integrates cardiovascular-specific fine-tuning, prompt-based inference, and entity-aware reasoning. Evaluations on MIMIC-III and CARDIO-NLP datasets demonstrate improved performance in precision, recall, F1-score, and AUROC, with high clinical relevance (kappa = 0.82) assessed by cardiologists. Challenges such as contextual hallucination, which occurs when plausible information contracts with provided source, and temporal ambiguity, which is related with models struggling with chronological ordering of events are addressed using prompt engineering and hybrid rule-based verification. This work underscores the potential of LLMs in clinical decision support systems (CDSS), advancing early warning systems and enhancing the translation of patient narratives into actionable risk assessments.
☆ Journalism-Guided Agentic In-Context Learning for News Stance Detection
As online news consumption grows, personalized recommendation systems have become integral to digital journalism. However, these systems risk reinforcing filter bubbles and political polarization by failing to incorporate diverse perspectives. Stance detection -- identifying a text's position on a target -- can help mitigate this by enabling viewpoint-aware recommendations and data-driven analyses of media bias. Yet, existing stance detection research remains largely limited to short texts and high-resource languages. To address these gaps, we introduce \textsc{K-News-Stance}, the first Korean dataset for article-level stance detection, comprising 2,000 news articles with article-level and 19,650 segment-level stance annotations across 47 societal issues. We also propose \textsc{JoA-ICL}, a \textbf{Jo}urnalism-guided \textbf{A}gentic \textbf{I}n-\textbf{C}ontext \textbf{L}earning framework that employs a language model agent to predict the stances of key structural segments (e.g., leads, quotes), which are then aggregated to infer the overall article stance. Experiments show that \textsc{JoA-ICL} outperforms existing stance detection methods, highlighting the benefits of segment-level agency in capturing the overall position of long-form news articles. Two case studies further demonstrate its broader utility in promoting viewpoint diversity in news recommendations and uncovering patterns of media bias.
comment: Preprint. 24 pages
☆ First-Order Error Matters: Accurate Compensation for Quantized Large Language Models
Post-training quantization (PTQ) offers an efficient approach to compressing large language models (LLMs), significantly reducing memory access and computational costs. Existing compensation-based weight calibration methods often rely on a second-order Taylor expansion to model quantization error, under the assumption that the first-order term is negligible in well-trained full-precision models. However, we reveal that the progressive compensation process introduces accumulated first-order deviations between latent weights and their full-precision counterparts, making this assumption fundamentally flawed. To address this, we propose FOEM, a novel PTQ method that explicitly incorporates first-order gradient terms to improve quantization error compensation. FOEM approximates gradients by directly computing the difference between latent and full-precision weights, avoiding the high cost and limited generalization of backpropagation-based gradient computation. This approach introduces minimal additional computational overhead. Moreover, FOEM leverages precomputed Cholesky factors to efficiently recover the inverse of Hessian submatrices in real time. Extensive experiments across a wide range of models and benchmarks demonstrate that FOEM consistently outperforms the classical GPTQ method. In 3-bit weight-only quantization, FOEM reduces the perplexity of Llama3-8B by 89.6%, and improves the 5-shot MMLU accuracy of Llama3-70B from 51.7% to 74.9%, approaching the full-precision performance of 78.6%. Furthermore, FOEM can be seamlessly integrated with advanced techniques such as GPTAQ and SpinQuant, yielding additional improvements under the challenging W4A4KV4 setting, and further narrowing the accuracy gap with full-precision baselines beyond what current state-of-the-art methods achieve. The code is available at https://github.com/Xingyu-Zheng/FOEM.
☆ Team HUMANE at AVeriTeC 2025: HerO 2 for Efficient Fact Verification ACL 2025
This paper presents HerO 2, Team HUMANE's system for the AVeriTeC shared task at the FEVER-25 workshop. HerO 2 is an enhanced version of HerO, the best-performing open-source model from the previous year's challenge. It improves evidence quality through document summarization and answer reformulation, optimizes veracity prediction via post-training quantization under computational constraints, and enhances overall system performance by integrating updated language model (LM) backbones. HerO 2 ranked second on the leaderboard while achieving the shortest runtime among the top three systems, demonstrating both high efficiency and strong potential for real-world fact verification. The code is available at https://github.com/ssu-humane/HerO2.
comment: ACL 2025 Workshop (FEVER)
☆ Mario at EXIST 2025: A Simple Gateway to Effective Multilingual Sexism Detection
This paper presents our approach to EXIST 2025 Task 1, addressing text-based sexism detection in English and Spanish tweets through hierarchical Low-Rank Adaptation (LoRA) of Llama 3.1 8B. Our method introduces conditional adapter routing that explicitly models label dependencies across three hierarchically structured subtasks: binary sexism identification, source intention detection, and multilabel sexism categorization. Unlike conventional LoRA applications that target only attention layers, we apply adaptation to all linear transformations, enhancing the model's capacity to capture task-specific patterns. In contrast to complex data processing and ensemble approaches, we show that straightforward parameter-efficient fine-tuning achieves strong performance. We train separate LoRA adapters (rank=16, QLoRA 4-bit) for each subtask using unified multilingual training that leverages Llama 3.1's native bilingual capabilities. The method requires minimal preprocessing and uses standard supervised learning. Our multilingual training strategy eliminates the need for separate language-specific models, achieving 1.7-2.4\% F1 improvements through cross-lingual transfer. With only 1.67\% trainable parameters compared to full fine-tuning, our approach reduces training time by 75\% and model storage by 98\%, while achieving competitive performance across all subtasks (ICM-Hard: 0.6774 for binary classification, 0.4991 for intention detection, 0.6519 for multilabel categorization).
comment: 12 pages, 5 tables, CLEF 2025
☆ Teach Me Sign: Stepwise Prompting LLM for Sign Language Production ICIP 2025
Large language models, with their strong reasoning ability and rich knowledge, have brought revolution to many tasks of AI, but their impact on sign language generation remains limited due to its complexity and unique rules. In this paper, we propose TEAch Me Sign (TEAM-Sign), treating sign language as another natural language. By fine-tuning an LLM, we enable it to learn the correspondence between text and sign language, and facilitate generation. Considering the differences between sign and spoken language, we employ a stepwise prompting strategy to extract the inherent sign language knowledge within the LLM, thereby supporting the learning and generation process. Experimental results on How2Sign and Phoenix14T datasets demonstrate that our approach effectively leverages both the sign language knowledge and reasoning capabilities of LLM to align the different distribution and grammatical rules between sign and spoken language.
comment: Accepted by IEEE ICIP 2025
☆ DS@GT at eRisk 2025: From prompts to predictions, benchmarking early depression detection with conversational agent based assessments and temporal attention models
This Working Note summarizes the participation of the DS@GT team in two eRisk 2025 challenges. For the Pilot Task on conversational depression detection with large language-models (LLMs), we adopted a prompt-engineering strategy in which diverse LLMs conducted BDI-II-based assessments and produced structured JSON outputs. Because ground-truth labels were unavailable, we evaluated cross-model agreement and internal consistency. Our prompt design methodology aligned model outputs with BDI-II criteria and enabled the analysis of conversational cues that influenced the prediction of symptoms. Our best submission, second on the official leaderboard, achieved DCHR = 0.50, ADODL = 0.89, and ASHR = 0.27.
☆ Modeling Understanding of Story-Based Analogies Using Large Language Models
Recent advancements in Large Language Models (LLMs) have brought them closer to matching human cognition across a variety of tasks. How well do these models align with human performance in detecting and mapping analogies? Prior research has shown that LLMs can extract similarities from analogy problems but lack robust human-like reasoning. Building on Webb, Holyoak, and Lu (2023), the current study focused on a story-based analogical mapping task and conducted a fine-grained evaluation of LLM reasoning abilities compared to human performance. First, it explored the semantic representation of analogies in LLMs, using sentence embeddings to assess whether they capture the similarity between the source and target texts of an analogy, and the dissimilarity between the source and distractor texts. Second, it investigated the effectiveness of explicitly prompting LLMs to explain analogies. Throughout, we examine whether LLMs exhibit similar performance profiles to those observed in humans by evaluating their reasoning at the level of individual analogies, and not just at the level of overall accuracy (as prior studies have done). Our experiments include evaluating the impact of model size (8B vs. 70B parameters) and performance variation across state-of-the-art model architectures such as GPT-4 and LLaMA3. This work advances our understanding of the analogical reasoning abilities of LLMs and their potential as models of human reasoning.
comment: To appear at CogSci 2025
☆ HanjaBridge: Resolving Semantic Ambiguity in Korean LLMs via Hanja-Augmented Pre-Training
Large language models (LLMs) often show poor performance in low-resource languages like Korean, partly due to unique linguistic challenges such as homophonous Sino-Korean words that are indistinguishable in Hangul script. To address this semantic ambiguity, we propose HanjaBridge, a novel meaning-injection technique integrated into a continual pre-training (CPT) framework. Instead of deterministically mapping a word to a single Hanja (Chinese character), HanjaBridge presents the model with all possible Hanja candidates for a given homograph, encouraging the model to learn contextual disambiguation. This process is paired with token-level knowledge distillation to prevent catastrophic forgetting. Experimental results show that HanjaBridge significantly improves Korean language understanding, achieving a 21\% relative improvement on the KoBALT benchmark. Notably, by reinforcing semantic alignment between Korean and Chinese through shared Hanja, we observe a strong positive cross-lingual transfer. Furthermore, these gains persist even when Hanja augmentation is omitted at inference time, ensuring practical efficiency with no additional run-time cost.
☆ How Stylistic Similarity Shapes Preferences in Dialogue Dataset with User and Third Party Evaluations SIGDIAL 2025
Recent advancements in dialogue generation have broadened the scope of human-bot interactions, enabling not only contextually appropriate responses but also the analysis of human affect and sensitivity. While prior work has suggested that stylistic similarity between user and system may enhance user impressions, the distinction between subjective and objective similarity is often overlooked. To investigate this issue, we introduce a novel dataset that includes users' preferences, subjective stylistic similarity based on users' own perceptions, and objective stylistic similarity annotated by third party evaluators in open-domain dialogue settings. Analysis using the constructed dataset reveals a strong positive correlation between subjective stylistic similarity and user preference. Furthermore, our analysis suggests an important finding: users' subjective stylistic similarity differs from third party objective similarity. This underscores the importance of distinguishing between subjective and objective evaluations and understanding the distinct aspects each captures when analyzing the relationship between stylistic similarity and user preferences. The dataset presented in this paper is available online.
comment: Accepted to SIGDIAL 2025 (long)
☆ LiLM-RDB-SFC: Lightweight Language Model with Relational Database-Guided DRL for Optimized SFC Provisioning
Effective management of Service Function Chains (SFCs) and optimal Virtual Network Function (VNF) placement are critical challenges in modern Software-Defined Networking (SDN) and Network Function Virtualization (NFV) environments. Although Deep Reinforcement Learning (DRL) is widely adopted for dynamic network decision-making, its inherent dependency on structured data and fixed action rules often limits adaptability and responsiveness, particularly under unpredictable network conditions. This paper introduces LiLM-RDB-SFC, a novel approach combining Lightweight Language Model (LiLM) with Relational Database (RDB) to answer network state queries to guide DRL model for efficient SFC provisioning. Our proposed approach leverages two LiLMs, Bidirectional and Auto-Regressive Transformers (BART) and the Fine-tuned Language Net T5 (FLAN-T5), to interpret network data and support diverse query types related to SFC demands, data center resources, and VNF availability. Results demonstrate that FLAN-T5 outperforms BART with a lower test loss (0.00161 compared to 0.00734), higher accuracy (94.79% compared to 80.2%), and less processing time (2h 2min compared to 2h 38min). Moreover, when compared to the large language model SQLCoder, FLAN-T5 matches the accuracy of SQLCoder while cutting processing time by 96% (SQLCoder: 54 h 43 min; FLAN-T5: 2 h 2 min).
comment: 9 pages, 6 figures, Accepted to IEEE 16th International Conference on Network of the Future (NoF) 2025
☆ NavComposer: Composing Language Instructions for Navigation Trajectories through Action-Scene-Object Modularization
Language-guided navigation is a cornerstone of embodied AI, enabling agents to interpret language instructions and navigate complex environments. However, expert-provided instructions are limited in quantity, while synthesized annotations often lack quality, making them insufficient for large-scale research. To address this, we propose NavComposer, a novel framework for automatically generating high-quality navigation instructions. NavComposer explicitly decomposes semantic entities such as actions, scenes, and objects, and recomposes them into natural language instructions. Its modular architecture allows flexible integration of state-of-the-art techniques, while the explicit use of semantic entities enhances both the richness and accuracy of instructions. Moreover, it operates in a data-agnostic manner, supporting adaptation to diverse navigation trajectories without domain-specific training. Complementing NavComposer, we introduce NavInstrCritic, a comprehensive annotation-free evaluation system that assesses navigation instructions on three dimensions: contrastive matching, semantic consistency, and linguistic diversity. NavInstrCritic provides a holistic evaluation of instruction quality, addressing limitations of traditional metrics that rely heavily on expert annotations. By decoupling instruction generation and evaluation from specific navigation agents, our method enables more scalable and generalizable research. Extensive experiments provide direct and practical evidence for the effectiveness of our method.
☆ Domain-Adaptive Small Language Models for Structured Tax Code Prediction
Every day, multinational firms process thousands of transactions, each of which must adhere to tax regulations that vary by jurisdiction and are often nuanced. The determination of product and service tax codes, such as HSN or SAC is a major use case in Tax compliance. An accurate determination of such codes is imperative to avoid any tax penalties. This paper proposes a domain-adaptive small language model (SLM) with an encoder-decoder architecture for the enhanced prediction of product and service tax codes. In this approach, we address the problem of predicting hierarchical tax code sequences using unstructured product and services data. We employ an SLM based upon encoder-decoder architecture as this enables sequential generation of tax codes to capture the hierarchical dependencies present within the tax codes. Our experiments demonstrate that encoder-decoder SLMs can be successfully applied to the sequential prediction of structured tax codes, a domain that remains comparatively unexplored in current NLP research. In this paper, we demonstrate the superior performance of the domain-adaptive encoder-decoder SLMs over flat classifiers when applied to the Harmonized System of Nomenclature (HSN), and achieve superior results compared to decoder-only and encoder-only architectures for structured sequence generation tasks. This approach can also be scaled to other government-mandated tax commodity codes, such as United Nations Standard Products and Services Codes (UNSPSC), or Brazil's Nomenclatura Comum do Mercosul (NCM).
comment: 10 pages, 3 figures
☆ Simulated Language Acquisition in a Biologically Realistic Model of the Brain
Despite tremendous progress in neuroscience, we do not have a compelling narrative for the precise way whereby the spiking of neurons in our brain results in high-level cognitive phenomena such as planning and language. We introduce a simple mathematical formulation of six basic and broadly accepted principles of neuroscience: excitatory neurons, brain areas, random synapses, Hebbian plasticity, local inhibition, and inter-area inhibition. We implement a simulated neuromorphic system based on this formalism, which is capable of basic language acquisition: Starting from a tabula rasa, the system learns, in any language, the semantics of words, their syntactic role (verb versus noun), and the word order of the language, including the ability to generate novel sentences, through the exposure to a modest number of grounded sentences in the same language. We discuss several possible extensions and implications of this result.
comment: 13 pages, 6 figures
☆ AI Wizards at CheckThat! 2025: Enhancing Transformer-Based Embeddings with Sentiment for Subjectivity Detection in News Articles
This paper presents AI Wizards' participation in the CLEF 2025 CheckThat! Lab Task 1: Subjectivity Detection in News Articles, classifying sentences as subjective/objective in monolingual, multilingual, and zero-shot settings. Training/development datasets were provided for Arabic, German, English, Italian, and Bulgarian; final evaluation included additional unseen languages (e.g., Greek, Romanian, Polish, Ukrainian) to assess generalization. Our primary strategy enhanced transformer-based classifiers by integrating sentiment scores, derived from an auxiliary model, with sentence representations, aiming to improve upon standard fine-tuning. We explored this sentiment-augmented architecture with mDeBERTaV3-base, ModernBERT-base (English), and Llama3.2-1B. To address class imbalance, prevalent across languages, we employed decision threshold calibration optimized on the development set. Our experiments show sentiment feature integration significantly boosts performance, especially subjective F1 score. This framework led to high rankings, notably 1st for Greek (Macro F1 = 0.51).
comment: 14 pages, 6 figures, accepted at CLEF 2025 CheckThat! Lab
☆ CRABS: A syntactic-semantic pincer strategy for bounding LLM interpretation of Python notebooks
Recognizing the information flows and operations comprising data science and machine learning Python notebooks is critical for evaluating, reusing, and adapting notebooks for new tasks. Investigating a notebook via re-execution often is impractical due to the challenges of resolving data and software dependencies. While Large Language Models (LLMs) pre-trained on large codebases have demonstrated effectiveness in understanding code without running it, we observe that they fail to understand some realistic notebooks due to hallucinations and long-context challenges. To address these issues, we propose a notebook understanding task yielding an information flow graph and corresponding cell execution dependency graph for a notebook, and demonstrate the effectiveness of a pincer strategy that uses limited syntactic analysis to assist full comprehension of the notebook using an LLM. Our Capture and Resolve Assisted Bounding Strategy (CRABS) employs shallow syntactic parsing and analysis of the abstract syntax tree (AST) to capture the correct interpretation of a notebook between lower and upper estimates of the inter-cell I/O sets, then uses an LLM to resolve remaining ambiguities via cell-by-cell zero-shot learning, thereby identifying the true data inputs and outputs of each cell. We evaluate and demonstrate the effectiveness of our approach using an annotated dataset of 50 representative, highly up-voted Kaggle notebooks that together represent 3454 actual cell inputs and outputs. The LLM correctly resolves 1397 of 1425 (98%) ambiguities left by analyzing the syntactic structure of these notebooks. Across 50 notebooks, CRABS achieves average F1 scores of 98% identifying cell-to-cell information flows and 99% identifying transitive cell execution dependencies.
comment: Preprint. Accepted to COLM 2025
☆ ExpliCIT-QA: Explainable Code-Based Image Table Question Answering
We present ExpliCIT-QA, a system that extends our previous MRT approach for tabular question answering into a multimodal pipeline capable of handling complex table images and providing explainable answers. ExpliCIT-QA follows a modular design, consisting of: (1) Multimodal Table Understanding, which uses a Chain-of-Thought approach to extract and transform content from table images; (2) Language-based Reasoning, where a step-by-step explanation in natural language is generated to solve the problem; (3) Automatic Code Generation, where Python/Pandas scripts are created based on the reasoning steps, with feedback for handling errors; (4) Code Execution to compute the final answer; and (5) Natural Language Explanation that describes how the answer was computed. The system is built for transparency and auditability: all intermediate outputs, parsed tables, reasoning steps, generated code, and final answers are available for inspection. This strategy works towards closing the explainability gap in end-to-end TableVQA systems. We evaluated ExpliCIT-QA on the TableVQA-Bench benchmark, comparing it with existing baselines. We demonstrated improvements in interpretability and transparency, which open the door for applications in sensitive domains like finance and healthcare where auditing results are critical.
comment: This work has been accepted for presentation at the 24nd Portuguese Conference on Artificial Intelligence (EPIA 2025) and will be published in the proceedings by Springer in the Lecture Notes in Computer Science (LNCS) series. Please cite the published version when available
☆ MetaLint: Generalizable Idiomatic Code Quality Analysis through Instruction-Following and Easy-to-Hard Generalization
Large Language Models, though successful in code generation, struggle with code quality analysis because they are limited by static training data and can't easily adapt to evolving best practices. We introduce MetaLint, a new instruction-following framework that formulates code quality analysis as the task of detecting and fixing problematic semantic code fragments or code idioms based on high-level specifications. Unlike conventional approaches that train models on static, rule-based data, MetaLint employs instruction tuning on synthetic linter-generated data to support easy-to-hard generalization, enabling models to adapt to novel or complex code patterns without retraining. To evaluate this, we construct a benchmark of challenging idioms inspired by real-world coding standards such as Python Enhancement Proposals (PEPs) and assess whether MetaLint-trained models reason adaptively or simply memorize. Our results show that MetaLint improves generalization to unseen PEP idioms, achieving a 70.37% F-score on idiom detection with the highest recall (70.43%) among all evaluated models. It also achieves 26.73% on localization, competitive for its 4B parameter size and comparable to larger state-of-the-art models like o3-mini, highlighting its potential for future-proof code quality analysis.
☆ Let's Think in Two Steps: Mitigating Agreement Bias in MLLMs with Self-Grounded Verification
Verifiers -- functions assigning rewards to agent behavior -- have been key for AI progress in domains like math and board games. However, extending these gains to domains without clear-cut success criteria (e.g.,computer use) remains a challenge: while humans can recognize suitable outcomes, translating this intuition into scalable rules is non-trivial. Multimodal Large Language Models(MLLMs) emerge as a promising solution, given their world knowledge, human-preference alignment, and reasoning skills. We evaluate MLLMs as verifiers of agent trajectories across web navigation, computer use, and robotic manipulation, and identify a critical limitation: agreement bias, a strong tendency for MLLMs to favor information in their context window, often generating chains of thought to rationalize flawed behavior. This bias is pervasive across models, resilient to test-time scaling, and can impact several methods using MLLMs as evaluators (e.g.,data filtering). Notably, it occurs despite MLLMs showing strong, human-aligned priors on desired behavior. To address this, we propose Self-Grounded Verification (SGV), a lightweight method that enables more effective use of MLLMs' knowledge and reasoning by harnessing their own sampling mechanisms via unconditional and conditional generation. SGV operates in two steps: first, the MLLM is elicited to retrieve broad priors about task completion, independent of the data under evaluation. Then, conditioned on self-generated priors, it reasons over and evaluates a candidate trajectory. Enhanced with SGV, MLLM verifiers show gains of up to 20 points in accuracy and failure detection rates, and can perform real-time supervision of heterogeneous agents, boosting task completion of a GUI specialist in OSWorld, a diffusion policy in robomimic, and a ReAct agent in VisualWebArena -- setting a new state of the art on the benchmark, surpassing the previous best by 48%.
comment: Our code and data are publicly available at https://github.com/mshalimay/mllm-verifiers-abias-sgv
☆ Partitioner Guided Modal Learning Framework
Multimodal learning benefits from multiple modal information, and each learned modal representations can be divided into uni-modal that can be learned from uni-modal training and paired-modal features that can be learned from cross-modal interaction. Building on this perspective, we propose a partitioner-guided modal learning framework, PgM, which consists of the modal partitioner, uni-modal learner, paired-modal learner, and uni-paired modal decoder. Modal partitioner segments the learned modal representation into uni-modal and paired-modal features. Modal learner incorporates two dedicated components for uni-modal and paired-modal learning. Uni-paired modal decoder reconstructs modal representation based on uni-modal and paired-modal features. PgM offers three key benefits: 1) thorough learning of uni-modal and paired-modal features, 2) flexible distribution adjustment for uni-modal and paired-modal representations to suit diverse downstream tasks, and 3) different learning rates across modalities and partitions. Extensive experiments demonstrate the effectiveness of PgM across four multimodal tasks and further highlight its transferability to existing models. Additionally, we visualize the distribution of uni-modal and paired-modal features across modalities and tasks, offering insights into their respective contributions.
comment: acm multimedia 2025
☆ Cross-lingual Few-shot Learning for Persian Sentiment Analysis with Incremental Adaptation
This research examines cross-lingual sentiment analysis using few-shot learning and incremental learning methods in Persian. The main objective is to develop a model capable of performing sentiment analysis in Persian using limited data, while getting prior knowledge from high-resource languages. To achieve this, three pre-trained multilingual models (XLM-RoBERTa, mDeBERTa, and DistilBERT) were employed, which were fine-tuned using few-shot and incremental learning approaches on small samples of Persian data from diverse sources, including X, Instagram, Digikala, Snappfood, and Taaghche. This variety enabled the models to learn from a broad range of contexts. Experimental results show that the mDeBERTa and XLM-RoBERTa achieved high performances, reaching 96% accuracy on Persian sentiment analysis. These findings highlight the effectiveness of combining few-shot learning and incremental learning with multilingual pre-trained models.
comment: Proceedings of the First National Conference on Artificial Intelligence and Emerging Research: Convergence of Humans and Intelligent Systems
☆ Jailbreak-Tuning: Models Efficiently Learn Jailbreak Susceptibility
AI systems are rapidly advancing in capability, and frontier model developers broadly acknowledge the need for safeguards against serious misuse. However, this paper demonstrates that fine-tuning, whether via open weights or closed fine-tuning APIs, can produce helpful-only models. In contrast to prior work which is blocked by modern moderation systems or achieved only partial removal of safeguards or degraded output quality, our jailbreak-tuning method teaches models to generate detailed, high-quality responses to arbitrary harmful requests. For example, OpenAI, Google, and Anthropic models will fully comply with requests for CBRN assistance, executing cyberattacks, and other criminal activity. We further show that backdoors can increase not only the stealth but also the severity of attacks, while stronger jailbreak prompts become even more effective in fine-tuning attacks, linking attack and potentially defenses in the input and weight spaces. Not only are these models vulnerable, more recent ones also appear to be becoming even more vulnerable to these attacks, underscoring the urgent need for tamper-resistant safeguards. Until such safeguards are discovered, companies and policymakers should view the release of any fine-tunable model as simultaneously releasing its evil twin: equally capable as the original model, and usable for any malicious purpose within its capabilities.
☆ MapIQ: Benchmarking Multimodal Large Language Models for Map Question Answering
Recent advancements in multimodal large language models (MLLMs) have driven researchers to explore how well these models read data visualizations, e.g., bar charts, scatter plots. More recently, attention has shifted to visual question answering with maps (Map-VQA). However, Map-VQA research has primarily focused on choropleth maps, which cover only a limited range of thematic categories and visual analytical tasks. To address these gaps, we introduce MapIQ, a benchmark dataset comprising 14,706 question-answer pairs across three map types: choropleth maps, cartograms, and proportional symbol maps spanning topics from six distinct themes (e.g., housing, crime). We evaluate multiple MLLMs using six visual analytical tasks, comparing their performance against one another and a human baseline. An additional experiment examining the impact of map design changes (e.g., altered color schemes, modified legend designs, and removal of map elements) provides insights into the robustness and sensitivity of MLLMs, their reliance on internal geographic knowledge, and potential avenues for improving Map-VQA performance.
comment: Published as a conference paper at COLM 2025
☆ LRCTI: A Large Language Model-Based Framework for Multi-Step Evidence Retrieval and Reasoning in Cyber Threat Intelligence Credibility Verification
Verifying the credibility of Cyber Threat Intelligence (CTI) is essential for reliable cybersecurity defense. However, traditional approaches typically treat this task as a static classification problem, relying on handcrafted features or isolated deep learning models. These methods often lack the robustness needed to handle incomplete, heterogeneous, or noisy intelligence, and they provide limited transparency in decision-making-factors that reduce their effectiveness in real-world threat environments. To address these limitations, we propose LRCTI, a Large Language Model (LLM)-based framework designed for multi-step CTI credibility verification. The framework first employs a text summarization module to distill complex intelligence reports into concise and actionable threat claims. It then uses an adaptive multi-step evidence retrieval mechanism that iteratively identifies and refines supporting information from a CTI-specific corpus, guided by LLM feedback. Finally, a prompt-based Natural Language Inference (NLI) module is applied to evaluate the credibility of each claim while generating interpretable justifications for the classification outcome. Experiments conducted on two benchmark datasets, CTI-200 and PolitiFact show that LRCTI improves F1-Macro and F1-Micro scores by over 5%, reaching 90.9% and 93.6%, respectively, compared to state-of-the-art baselines. These results demonstrate that LRCTI effectively addresses the core limitations of prior methods, offering a scalable, accurate, and explainable solution for automated CTI credibility verification
♻ ☆ LongDocURL: a Comprehensive Multimodal Long Document Benchmark Integrating Understanding, Reasoning, and Locating
Large vision language models (LVLMs) have improved the document understanding capabilities remarkably, enabling the handling of complex document elements, longer contexts, and a wider range of tasks. However, existing document understanding benchmarks have been limited to handling only a small number of pages and fail to provide a comprehensive analysis of layout elements locating. In this paper, we first define three primary task categories: Long Document Understanding, numerical Reasoning, and cross-element Locating, and then propose a comprehensive benchmark, LongDocURL, integrating above three primary tasks and comprising 20 sub-tasks categorized based on different primary tasks and answer evidences. Furthermore, we develop a semi-automated construction pipeline and collect 2,325 high-quality question-answering pairs, covering more than 33,000 pages of documents, significantly outperforming existing benchmarks. Subsequently, we conduct comprehensive evaluation experiments on both open-source and closed-source models across 26 different configurations, revealing critical performance gaps in this field.
♻ ☆ A Mathematical Theory of Discursive Networks
Large-language models (LLMs) turn writing into a live exchange between humans and software. We characterize this new medium as a discursive network that treats people and LLMs as equal nodes and tracks how their statements circulate. We define the generation of erroneous information as invalidation (any factual, logical, or structural breach) and show it follows four hazards: drift from truth, self-repair, fresh fabrication, and external detection. We develop a general mathematical model of discursive networks that shows that a network governed only by drift and self-repair stabilizes at a modest error rate. Giving each false claim even a small chance of peer review shifts the system to a truth-dominant state. We operationalize peer review with the open-source \emph{Flaws-of-Others (FOO) algorithm}: a configurable loop in which any set of agents critique one another while a harmonizer merges their verdicts. We identify an ethical transgression, epithesis, that occurs when humans fail to engage in the discursive network. The takeaway is practical and cultural: reliability in this new medium comes not from perfecting single models but from connecting imperfect ones into networks that enforce mutual accountability.
comment: 32 pages, 4 figures, 4 tables, 1 algorithm, 54 references
♻ ☆ Conversation Forests: The Key to Fine Tuning Large Language Models for Multi-Turn Medical Conversations is Branching
Fine-tuning methods such as Direct Preference Optimization (DPO) and Group Relative Policy Optimization (GRPO) have demonstrated success in training large language models (LLMs) for single-turn tasks. However, these methods fall short in multi-turn applications, such as diagnostic patient interviewing, where understanding how early conversational turns influence downstream completions and outcomes is essential. In medicine, a multi-turn perspective is critical for learning diagnostic schemas and better understanding conversation dynamics. To address this gap, I introduce Savage Conversation Forests (SCF), a reinforcement learning framework that leverages a branched conversation architecture to fine-tune LLMs for multi-turn dialogue. SCF generates multiple possible conversation continuations at each turn, enabling the model to learn how different early responses affect downstream interactions and diagnostic outcomes. In experiments simulating doctor-patient conversations, SCF with branching outperforms linear conversation architectures on diagnostic accuracy. I hypothesize that SCF's improvements stem from its ability to provide richer, interdependent training signals across conversation turns. These results suggest that a branched training architecture is an important strategy for fine tuning LLMs in complex multi-turn conversational tasks.
♻ ☆ ProtocolLLM: RTL Benchmark for SystemVerilog Generation of Communication Protocols ISCA 2025
Recent advances in large language models (LLMs) have demonstrated strong performance in generating code for general-purpose programming languages. However, their potential for hardware description languages (HDLs), such as SystemVerilog, remains largely unexplored. HDL code generation poses unique challenges due to strict timing semantics, concurrency, and synthesizability constraints essential for correct hardware functionality. Further, HDL-based design flows encompass a broad set of tasks beyond structural code generation, including testbench development, assertion-based verification, timing closure, and protocol-level integration for on-chip communication. In this work, we evaluate the capabilities of both open-source and state-of-the-art LLMs in generating synthesizable and functionally accurate SystemVerilog implementations of widely used communication protocols that are critical components of embedded and System-on-Chip (SoC) systems. We introduce ProtocolLLM, the first benchmark suite specifically targeting these protocols with tasks spanning multiple design abstraction levels and varying prompt specificity. Our evaluation method also focuses on timing correctness in addition to synthesizability and syntactic correctness. We observe that most of the models fail to generate SystemVerilog code for communication protocols that follow timing constrains.
comment: Accepted at MLSysArch@ISCA 2025
♻ ☆ A Generative Approach to LLM Harmfulness Detection with Special Red Flag Tokens
Most safety training methods for large language models (LLMs) are based on fine-tuning that forces models to shift from an unsafe answer to refusal when faced with harmful requests. Unfortunately, these drastic distribution shifts generally compromise model capabilities. To avoid that, we propose to expand the model's vocabulary with a special token we call red flag token () and propose to train the model to insert this token into its response at any time when harmful content is generated or about to be generated. Our approach offers several advantages: it enables the model to explicitly learn the concept of harmfulness while marginally affecting the generated distribution, thus maintaining the model's utility. It also evaluates each generated answer and provides robustness as good as adversarial training without the need to run attacks during training. Moreover, by encapsulating our safety tuning in a LoRA module, we provide additional defenses against fine-tuning API attacks.
comment: 14 pages, 6 figures
♻ ☆ Hallucination Stations: On Some Basic Limitations of Transformer-Based Language Models AAAI-26
In this paper we explore hallucinations and related capability limitations in LLMs and LLM-based agents from the perspective of computational complexity. We show that beyond a certain complexity, LLMs are incapable of carrying out computational and agentic tasks or verifying their accuracy.
comment: 6 pages; to be submitted to AAAI-26 after reviews
♻ ☆ Gaussian mixture models as a proxy for interacting language models
Large language models (LLMs) are a powerful tool with the ability to match human capabilities and behavior in many settings. Retrieval-augmented generation (RAG) further allows LLMs to generate diverse output depending on the contents of their RAG database. This motivates their use in the social sciences to study human behavior between individuals when large-scale experiments are infeasible. However, LLMs depend on complex, computationally expensive algorithms. In this paper, we introduce interacting Gaussian mixture models (GMMs) as an alternative to similar frameworks using LLMs. We compare a simplified model of GMMs to select experimental simulations of LLMs whose updating and response depend on feedback from other LLMs. We find that interacting GMMs capture important features of the dynamics in interacting LLMs, and we investigate key similarities and differences between interacting LLMs and GMMs. We conclude by discussing the benefits of Gaussian mixture models, potential modifications, and future research directions.
♻ ☆ Following the Clues: Experiments on Person Re-ID using Cross-Modal Intelligence SC 2025
The collection and release of street-level recordings as Open Data play a vital role in advancing autonomous driving systems and AI research. However, these datasets pose significant privacy risks, particularly for pedestrians, due to the presence of Personally Identifiable Information (PII) that extends beyond biometric traits such as faces. In this paper, we present cRID, a novel cross-modal framework combining Large Vision-Language Models, Graph Attention Networks, and representation learning to detect textual describable clues of PII and enhance person re-identification (Re-ID). Our approach focuses on identifying and leveraging interpretable features, enabling the detection of semantically meaningful PII beyond low-level appearance cues. We conduct a systematic evaluation of PII presence in person image datasets. Our experiments show improved performance in practical cross-dataset Re-ID scenarios, notably from Market-1501 to CUHK03-np (detected), highlighting the framework's practical utility. Code is available at https://github.com/RAufschlaeger/cRID.
comment: accepted for publication at the 2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC 2025), taking place during November 18-21, 2025 in Gold Coast, Australia
♻ ☆ Truthful or Fabricated? Using Causal Attribution to Mitigate Reward Hacking in Explanations
Chain-of-thought explanations are widely used to inspect the decision process of large language models (LLMs) and to evaluate the trustworthiness of model outputs, making them important for effective collaboration between LLMs and humans. We demonstrate that preference optimization - a key step in the alignment phase - can inadvertently reduce the faithfulness of these explanations. This occurs because the reward model (RM), which guides alignment, is tasked with optimizing both the expected quality of the response and the appropriateness of the explanations (e.g., minimizing bias or adhering to safety standards), creating potential conflicts. The RM lacks a mechanism to assess the consistency between the model's internal decision process and the generated explanation. Consequently, the LLM may engage in "reward hacking" by producing a final response that scores highly while giving an explanation tailored to maximize reward rather than accurately reflecting its reasoning. To address this issue, we propose enriching the RM's input with a causal attribution of the prediction, allowing the RM to detect discrepancies between the generated self-explanation and the model's decision process. In controlled settings, we show that this approach reduces the tendency of the LLM to generate misleading explanations.
comment: 20 pages, 10 figures, 6 tables
♻ ☆ ETT: Expanding the Long Context Understanding Capability of LLMs at Test-Time
Transformer-based Language Models' computation and memory overhead increase quadratically as a function of sequence length. The quadratic cost poses challenges when employing LLMs for processing long sequences. In this work, we introduce \ourmodelacronym~(Extend at Test-Time), method for extending the context length of short context Transformer-based LLMs, with constant memory requirement and linear computation overhead. ETT enable the extension of the context length at test-time by efficient fine-tuning the model's parameters on the input context, chunked into overlapping small subsequences. We evaluate ETT on LongBench by extending the context length of GPT-Large and Phi-2 up to 32 times, increasing from 1k to 32k tokens. This results in up to a 30 percent improvement in the model's accuracy. We also study how context can be stored in LLM's weights effectively and efficiently. Through a detailed ablation study, we examine which Transformer modules are most beneficial to fine-tune at test-time. Interestingly, we find that fine-tuning the second layer of the FFNs is more effective than full fine-tuning, leading to a further improvement in the models' accuracy.
♻ ☆ ImpliRet: Benchmarking the Implicit Fact Retrieval Challenge
Retrieval systems are central to many NLP pipelines, but often rely on surface-level cues such as keyword overlap and lexical semantic similarity. To evaluate retrieval beyond these shallow signals, recent benchmarks introduce reasoning-heavy queries; however, they primarily shift the burden to query-side processing techniques -- like prompting or multi-hop retrieval -- that can help resolve complexity. In contrast, we present ImpliRet, a benchmark that shifts the reasoning challenge to document-side processing: The queries are simple, but relevance depends on facts stated implicitly in documents through temporal (e.g., resolving "two days ago"), arithmetic, and world knowledge relationships. We evaluate a range of sparse and dense retrievers, all of which struggle in this setting: the best nDCG@10 is only 14.91%. We also test whether long-context models can overcome this limitation. But even with a short context of only thirty documents, including the positive document, GPT-o4-mini scores only 55.54%, showing that document-side reasoning remains a challenge. Our codes are available at: github.com/ZeinabTaghavi/IMPLIRET
♻ ☆ ContextCache: Context-Aware Semantic Cache for Multi-Turn Queries in Large Language Models
Semantic caching significantly reduces computational costs and improves efficiency by storing and reusing large language model (LLM) responses. However, existing systems rely primarily on matching individual queries, lacking awareness of multi-turn dialogue contexts, which leads to incorrect cache hits when similar queries appear in different conversational settings. This demonstration introduces ContextCache, a context-aware semantic caching system for multi-turn dialogues. ContextCache employs a two-stage retrieval architecture that first executes vector-based retrieval on the current query to identify potential matches and then integrates current and historical dialogue representations through self-attention mechanisms for precise contextual matching. Evaluation of real-world conversations shows that ContextCache improves precision and recall compared to existing methods. Additionally, cached responses exhibit approximately 10 times lower latency than direct LLM invocation, enabling significant computational cost reductions for LLM conversational applications.
♻ ☆ Block Circulant Adapter for Large Language Models IJCAI-2025
Fine-tuning large language models (LLMs) is difficult due to their huge model size. Recent Fourier domain-based methods show potential for reducing fine-tuning costs. We propose a block circulant matrix-based fine-tuning method with a stable training heuristic to leverage the properties of circulant matrices and one-dimensional Fourier transforms to reduce storage and computation costs. Experiments show that our method uses $14\times$ less number of parameters than VeRA, $16\times$ smaller than LoRA and $32\times$ less FLOPs than FourierFT, while maintaining close or better task performance. Our approach presents a promising way in frequency domain to fine-tune large models on downstream tasks.
comment: to appear in Proceedings of the 2025 International Joint Conference on Artificial Intelligence (IJCAI-2025)
♻ ☆ Shared Global and Local Geometry of Language Model Embeddings
Researchers have recently suggested that models share common representations. In our work, we find numerous geometric similarities across the token embeddings of large language models. First, we find ``global'' similarities: token embeddings often share similar relative orientations. Next, we characterize local geometry in two ways: (1) by using Locally Linear Embeddings, and (2) by defining a simple measure for the intrinsic dimension of each embedding. Both characterizations allow us to find local similarities across token embeddings. Additionally, our intrinsic dimension demonstrates that embeddings lie on a lower dimensional manifold, and that tokens with lower intrinsic dimensions often have semantically coherent clusters, while those with higher intrinsic dimensions do not. Based on our findings, we introduce EMB2EMB, a simple application to linearly transform steering vectors from one language model to another, despite the two models having different dimensions.
♻ ☆ Critique-GRPO: Advancing LLM Reasoning with Natural Language and Numerical Feedback
Recent advances in reinforcement learning (RL) with numerical feedback, such as scalar rewards, have significantly enhanced the complex reasoning capabilities of large language models (LLMs). Despite this success, we identify three key challenges encountered by RL with solely numerical feedback: performance plateaus, limited effectiveness of self-reflection, and persistent failures. We then demonstrate that RL-finetuned models, even after exhibiting performance plateaus, can generate correct refinements on persistently failed problems by leveraging natural language feedback in the form of critiques. Building on this insight, we propose Critique-GRPO, an online RL framework that integrates both natural language and numerical feedback for effective policy optimization. Critique-GRPO enables LLMs to learn from initial responses and critique-guided self-refinements simultaneously while maintaining exploration. Additionally, we employ a shaping function to amplify learning from correct, especially unfamiliar, refinements and penalize incorrect ones. Extensive experiments with Qwen2.5-7B-Base, Qwen2.5-Math-7B-Base, and Qwen3-8B demonstrate that Critique-GRPO consistently outperforms supervised learning and RL-based fine-tuning methods across eight challenging mathematical, STEM, and general reasoning tasks, improving average pass@1 scores by approximately 4.4% and 3.8% on Qwen2.5-7B-Base and Qwen3-8B, respectively. Notably, Critique-GRPO enables effective self-improvement through self-critiquing and weak-to-strong generalization, achieving consistent gains over GRPO, such as 16.7% and 10.0% pass@1 improvements on AIME 2024, respectively.
comment: 49 pages, updated with new experimental results
♻ ☆ KAT-V1: Kwai-AutoThink Technical Report
We present Kwaipilot-AutoThink (KAT), an open-source 40B large language model developed to address the overthinking problem in reasoning-intensive tasks, where an automatic thinking training paradigm is proposed to dynamically switch between reasoning and non-reasoning modes based on task complexity. Specifically, first, we construct the dual-regime dataset based on a novel tagging pipeline and a multi-agent synthesis strategy, and then we apply Multi-Token Prediction (MTP)-enhanced knowledge distillation, enabling efficient and fine-grained reasoning transfer with minimal pretraining cost. Besides, we implement a cold-start initialization strategy that introduces mode-selection priors using majority-vote signals and intent-aware prompting. Finally, we propose Step-SRPO, a reinforcement learning algorithm that incorporates intermediate supervision into the GRPO framework, offering structured guidance over both reasoning-mode selection and response accuracy. Extensive experiments across multiple benchmarks demonstrate that KAT consistently matches or even outperforms current state-of-the-art models, including DeepSeek-R1-0528 and Qwen3-235B-A22B, across a wide range of reasoning-intensive tasks while reducing token usage by up to approximately 30\%. Beyond academic evaluation, KAT has been successfully deployed in Kwaipilot (i.e., Kuaishou's internal coding assistant), and improves real-world development workflows with high accuracy, efficiency, and controllable reasoning behaviors. Moreover, we are actively training a 200B Mixture-of-Experts (MoE) with 40B activation parameters, where the early-stage results already demonstrate promising improvements in performance and efficiency, further showing the scalability of the AutoThink paradigm.
♻ ☆ RAG-R1 : Incentivize the Search and Reasoning Capabilities of LLMs through Multi-query Parallelism
Large Language Models (LLMs) have demonstrated remarkable capabilities across various tasks, while they remain prone to generating hallucinated or outdated responses due to their static internal knowledge. Recent advancements in Retrieval-Augmented Generation (RAG) methods have explored enhancing models' search and reasoning capabilities through reinforcement learning (RL). Although these methods demonstrate promising results, they face challenges in training stability and encounter issues such as substantial inference time and restricted capabilities due to the single-query mode. In this paper, we propose RAG-R1, a novel training framework designed to enable LLMs to adaptively leverage internal and external knowledge during the reasoning process. We further expand the generation and retrieval processes within the framework from single-query mode to multi-query parallelism, aimed at reducing inference time and enhancing the model's capabilities. Extensive experiments on seven question-answering benchmarks demonstrate that our method outperforms the strongest baseline by up to 13.2% and decreases inference time by 11.1%.
♻ ☆ On the Effect of Instruction Tuning Loss on Generalization ACL
Instruction Tuning has emerged as a pivotal post-training paradigm that enables pre-trained language models to better follow user instructions. Despite its significance, little attention has been given to optimizing the loss function used. A fundamental, yet often overlooked, question is whether the conventional auto-regressive objective - where loss is computed only on response tokens, excluding prompt tokens - is truly optimal for instruction tuning. In this work, we systematically investigate the impact of differentially weighting prompt and response tokens in instruction tuning loss, and propose Weighted Instruction Tuning (WIT) as a better alternative to conventional instruction tuning. Through extensive experiments on five language models of different families and scale, three finetuning datasets of different sizes, and five diverse evaluation benchmarks, we show that the standard instruction tuning loss often yields suboptimal performance and limited robustness to input prompt variations. We find that a low-to-moderate weight for prompt tokens coupled with a moderate-to-high weight for response tokens yields the best-performing models across settings and also serve as better starting points for the subsequent preference alignment training. These findings highlight the need to reconsider instruction tuning loss and offer actionable insights for developing more robust and generalizable models. Our code is open-sourced at https://github.com/kowndinya-renduchintala/WIT.
comment: To appear in Transactions of the Association for Computational Linguistics (TACL)
♻ ☆ Stylometry recognizes human and LLM-generated texts in short samples
The paper explores stylometry as a method to distinguish between texts created by Large Language Models (LLMs) and humans, addressing issues of model attribution, intellectual property, and ethical AI use. Stylometry has been used extensively to characterise the style and attribute authorship of texts. By applying it to LLM-generated texts, we identify their emergent writing patterns. The paper involves creating a benchmark dataset based on Wikipedia, with (a) human-written term summaries, (b) texts generated purely by LLMs (GPT-3.5/4, LLaMa 2/3, Orca, and Falcon), (c) processed through multiple text summarisation methods (T5, BART, Gensim, and Sumy), and (d) rephrasing methods (Dipper, T5). The 10-sentence long texts were classified by tree-based models (decision trees and LightGBM) using human-designed (StyloMetrix) and n-gram-based (our own pipeline) stylometric features that encode lexical, grammatical, syntactic, and punctuation patterns. The cross-validated results reached a performance of up to .87 Matthews correlation coefficient in the multiclass scenario with 7 classes, and accuracy between .79 and 1. in binary classification, with the particular example of Wikipedia and GPT-4 reaching up to .98 accuracy on a balanced dataset. Shapley Additive Explanations pinpointed features characteristic of the encyclopaedic text type, individual overused words, as well as a greater grammatical standardisation of LLMs with respect to human-written texts. These results show -- crucially, in the context of the increasingly sophisticated LLMs -- that it is possible to distinguish machine- from human-generated texts at least for a well-defined text type.
♻ ☆ SocioVerse: A World Model for Social Simulation Powered by LLM Agents and A Pool of 10 Million Real-World Users
Social simulation is transforming traditional social science research by modeling human behavior through interactions between virtual individuals and their environments. With recent advances in large language models (LLMs), this approach has shown growing potential in capturing individual differences and predicting group behaviors. However, existing methods face alignment challenges related to the environment, target users, interaction mechanisms, and behavioral patterns. To this end, we introduce SocioVerse, an LLM-agent-driven world model for social simulation. Our framework features four powerful alignment components and a user pool of 10 million real individuals. To validate its effectiveness, we conducted large-scale simulation experiments across three distinct domains: politics, news, and economics. Results demonstrate that SocioVerse can reflect large-scale population dynamics while ensuring diversity, credibility, and representativeness through standardized procedures and minimal manual adjustments.
♻ ☆ Text to model via SysML: Automated generation of dynamical system computational models from unstructured natural language text via enhanced System Modeling Language diagrams
This paper contributes to speeding up the design and deployment of engineering dynamical systems by proposing a strategy for exploiting domain and expert knowledge for the automated generation of dynamical system computational model starting from a corpus of document relevant to the dynamical system of interest and an input document describing the specific system. This strategy is implemented in five steps and, crucially, it uses system modeling language diagrams (SysML) to extract accurate information about the dependencies, attributes, and operations of components. Natural Language Processing (NLP) strategies and Large Language Models (LLMs) are employed in specific tasks to improve intermediate outputs of the SySML diagrams automated generation, such as: list of key nouns; list of extracted relationships; list of key phrases and key relationships; block attribute values; block relationships; and BDD diagram generation. The applicability of automated SysML diagram generation is illustrated with different case studies. The computational models of complex dynamical systems from SysML diagrams are then obtained via code generation and computational model generation steps. In the code generation step, NLP strategies are used for summarization, while LLMs are used for validation only. The proposed approach is not limited to a specific system, domain, or computational software. The applicability of the proposed approach is shown via an end-to-end example from text to model of a simple pendulum, showing improved performance compared to results yielded by LLMs only.
comment: v2 - typos and imprecisions corrected
♻ ☆ Compression Hacking: A Supplementary Perspective on Informatics Properties of Language Models from Geometric Distortion
Recently, the concept of ``compression as intelligence'' has provided a novel informatics metric perspective for language models (LMs), emphasizing that highly structured representations signify the intelligence level of LMs. However, from a geometric standpoint, the word representation space of highly compressed LMs tends to degenerate into a highly anisotropic state, which hinders the LM's ability to comprehend instructions and directly impacts its performance. We found this compression-anisotropy synchronicity is essentially the ``Compression Hacking'' in LM representations, where noise-dominated directions tend to create the illusion of high compression rates by sacrificing spatial uniformity. Based on this, we propose three refined compression metrics by incorporating geometric distortion analysis and integrate them into a self-evaluation pipeline. The refined metrics exhibit strong alignment with the LM's comprehensive capabilities, achieving Spearman correlation coefficients above 0.9, significantly outperforming both the original compression and other internal structure-based metrics. This confirms that compression hacking substantially enhances the informatics interpretation of LMs by incorporating geometric distortion of representations.
♻ ☆ SECURE: Semantics-aware Embodied Conversation under Unawareness for Lifelong Robot Learning
This paper addresses a challenging interactive task learning scenario we call rearrangement under unawareness: an agent must manipulate a rigid-body environment without knowing a key concept necessary for solving the task and must learn about it during deployment. For example, the user may ask to "put the two granny smith apples inside the basket", but the agent cannot correctly identify which objects in the environment are "granny smith" as the agent has not been exposed to such a concept before. We introduce SECURE, an interactive task learning policy designed to tackle such scenarios. The unique feature of SECURE is its ability to enable agents to engage in semantic analysis when processing embodied conversations and making decisions. Through embodied conversation, a SECURE agent adjusts its deficient domain model by engaging in dialogue to identify and learn about previously unforeseen possibilities. The SECURE agent learns from the user's embodied corrective feedback when mistakes are made and strategically engages in dialogue to uncover useful information about novel concepts relevant to the task. These capabilities enable the SECURE agent to generalize to new tasks with the acquired knowledge. We demonstrate in the simulated Blocksworld and the real-world apple manipulation environments that the SECURE agent, which solves such rearrangements under unawareness, is more data-efficient than agents that do not engage in embodied conversation or semantic analysis.
comment: Published at 4th Conference on Lifelong Learning Agents (CoLLAs), 2025
♻ ☆ FalseReject: A Resource for Improving Contextual Safety and Mitigating Over-Refusals in LLMs via Structured Reasoning
Safety alignment approaches in large language models (LLMs) often lead to the over-refusal of benign queries, significantly diminishing their utility in sensitive scenarios. To address this challenge, we introduce FalseReject, a comprehensive resource containing 16k seemingly toxic queries accompanied by structured responses across 44 safety-related categories. We propose a graph-informed adversarial multi-agent interaction framework to generate diverse and complex prompts, while structuring responses with explicit reasoning to aid models in accurately distinguishing safe from unsafe contexts. FalseReject includes training datasets tailored for both standard instruction-tuned models and reasoning-oriented models, as well as a human-annotated benchmark test set. Our extensive benchmarking on 29 state-of-the-art (SOTA) LLMs reveals persistent over-refusal challenges. Empirical results demonstrate that supervised finetuning with FalseReject substantially reduces unnecessary refusals without compromising overall safety or general language capabilities.
comment: Accepted at COLM 2025
♻ ☆ Is Compression Really Linear with Code Intelligence?
Understanding the relationship between data compression and the capabilities of Large Language Models (LLMs) is crucial, especially in specialized domains like code intelligence. Prior work posited a linear relationship between compression and general intelligence. However, it overlooked the multifaceted nature of code that encompasses diverse programming languages and tasks, and struggled with fair evaluation of modern Code LLMs. We address this by evaluating a diverse array of open-source Code LLMs on comprehensive multi-language, multi-task code benchmarks. To address the challenge of efficient and fair evaluation of pre-trained LLMs' code intelligence, we introduce \textit{Format Annealing}, a lightweight, transparent training methodology designed to assess the intrinsic capabilities of these pre-trained models equitably. Compression efficacy, measured as bits-per-character (BPC), is determined using a novel, large-scale, and previously unseen code validation set derived from GitHub. Our empirical results reveal a fundamental logarithmic relationship between measured code intelligence and BPC. This finding refines prior hypotheses of linearity, which we suggest are likely observations of the logarithmic curve's tail under specific, limited conditions. Our work provides a more nuanced understanding of compression's role in developing code intelligence and contributes a robust evaluation framework in the code domain.
comment: work in progress
♻ ☆ Style over Substance: Distilled Language Models Reason Via Stylistic Replication
Specialized reasoning language models (RLMs) have demonstrated that scaling test-time computation through detailed reasoning traces significantly enhances performance. Although these traces effectively facilitate knowledge distillation into smaller, instruction-tuned models, the precise nature of transferred reasoning remains unclear. In this study, we investigate to what extent distilled models internalize replicated stylistic patterns during reasoning. To this end, we systematically analyze reasoning traces, identifying structural and lexical patterns that characterize successful reasoning. We then introduce two new datasets -- a dataset of emergent reasoning traces and a synthetic dataset explicitly constructed to replicate these stylistic patterns -- to precisely examine their influence on distilled models' reasoning capabilities. We find that models trained on the synthetic traces achieve comparable performance, indicating that distilled reasoning abilities rely significantly on surface-level patterns. Surprisingly, we observe an increase in performance even when the synthetic traces are altered to lead to the wrong answer. Our findings highlight how stylistic patterns can be leveraged to efficiently enhance LM reasoning across diverse model families.
comment: To appear at COLM 2025
♻ ☆ Plancraft: an evaluation dataset for planning with LLM agents
We present Plancraft, a multi-modal evaluation dataset for LLM agents. Plancraft has both a text-only and multi-modal interface, based on the Minecraft crafting GUI. We include the Minecraft Wiki to evaluate tool use and Retrieval Augmented Generation (RAG), as well as a handcrafted planner and Oracle Retriever, to ablate the different components of a modern agent architecture. To evaluate decision-making, Plancraft also includes a subset of examples that are intentionally unsolvable, providing a realistic challenge that requires the agent not only to complete tasks but also to decide whether they are solvable at all. We benchmark both open-source and closed-source LLMs and compare their performance and efficiency to a handcrafted planner. Overall, we find that LLMs and VLMs struggle with the planning problems that Plancraft introduces, and offer suggestions on how to improve their capabilities.
♻ ☆ Evaluating Multimodal Large Language Models on Educational Textbook Question Answering
Multimodal large language models (MLLMs) have shown success in vision-language tasks, but their ability to reason over complex educational materials remains largely untested. This work presents the first evaluation of state-of-the-art MLLMs, including LLaVA-1.5 and LLaMA 3.2-Vision, on the textbook question answering (TQA) task using the CK12-QA dataset. We introduce a multimodal retrieval-augmented generation (RAG) pipeline to simulate real-world learning by providing relevant lesson paragraphs and diagrams as context. Our zero-shot experiments reveal a critical trade-off: while retrieved context improves LLaVA's performance on text-based questions, it significantly degrades the accuracy of the more powerful LLaMA 3.2-Vision on diagram-based tasks, dropping its validation accuracy from 74.07% to 25.93%. We term this statistically significant phenomenon "catastrophic context interference." Furthermore, fine-tuning highlights architectural differences: LLaMA 3.2-Vision's performance improves to 71.16% on the test set, demonstrating its capacity to learn multimodal integration, whereas LLaVA's performance declines, indicating challenges with generalization. Our results underscore the challenges MLLMs face in modality prioritization and context integration, providing a benchmark and pointing to key directions for developing more robust AI-driven educational tools.
comment: 8 Pages
♻ ☆ Traveling Across Languages: Benchmarking Cross-Lingual Consistency in Multimodal LLMs
The rapid evolution of multimodal large language models (MLLMs) has significantly enhanced their real-world applications. However, achieving consistent performance across languages, especially when integrating cultural knowledge, remains a significant challenge. To better assess this issue, we introduce two new benchmarks: KnowRecall and VisRecall, which evaluate cross-lingual consistency in MLLMs. KnowRecall is a visual question answering benchmark designed to measure factual knowledge consistency in 15 languages, focusing on cultural and historical questions about global landmarks. VisRecall assesses visual memory consistency by asking models to describe landmark appearances in 9 languages without access to images. Experimental results reveal that state-of-the-art MLLMs, including proprietary ones, still struggle to achieve cross-lingual consistency. This underscores the need for more robust approaches that produce truly multilingual and culturally aware models.
comment: https://github.com/nlp-waseda/traveling-across-languages
♻ ☆ Voting or Consensus? Decision-Making in Multi-Agent Debate ACL2025
Much of the success of multi-agent debates depends on carefully choosing the right parameters. The decision-making protocol stands out as it can highly impact final model answers, depending on how decisions are reached. Systematic comparison of decision protocols is difficult because many studies alter multiple discussion parameters beyond the protocol. So far, it has been largely unknown how decision-making influences different tasks. This work systematically evaluates the impact of seven decision protocols (e.g., majority voting, unanimity consensus). We change only one variable at a time - the decision protocol - to analyze how different methods affect the collaboration between agents and measure differences in knowledge and reasoning tasks. Our results show that voting protocols improve performance by 13.2% in reasoning tasks and consensus protocols by 2.8% in knowledge tasks compared to other decision protocols. Increasing the number of agents improves performance, while more discussion rounds before voting reduce it. To improve decision-making by increasing answer diversity, we propose two new methods, All-Agents Drafting (AAD) and Collective Improvement (CI). Our methods improve task performance by up to 3.3% with AAD and up to 7.4% with CI. This work demonstrates the importance of decision-making in multi-agent debates beyond scaling.
comment: Accepted at ACL2025 (Findings)
♻ ☆ Comply: Learning Sentences with Complex Weights inspired by Fruit Fly Olfaction
Biologically inspired neural networks offer alternative avenues to model data distributions. FlyVec is a recent example that draws inspiration from the fruit fly's olfactory circuit to tackle the task of learning word embeddings. Surprisingly, this model performs competitively even against deep learning approaches specifically designed to encode text, and it does so with the highest degree of computational efficiency. We pose the question of whether this performance can be improved further. For this, we introduce Comply. By incorporating positional information through complex weights, we enable a single-layer neural network to learn sequence representations. Our experiments show that Comply not only supersedes FlyVec but also performs on par with significantly larger state-of-the-art models. We achieve this without additional parameters. Comply yields sparse contextual representations of sentences that can be interpreted explicitly from the neuron weights.
comment: Accepted at NICE2025
♻ ☆ DocPolarBERT: A Pre-trained Model for Document Understanding with Relative Polar Coordinate Encoding of Layout Structures
We introduce DocPolarBERT, a layout-aware BERT model for document understanding that eliminates the need for absolute 2D positional embeddings. We extend self-attention to take into account text block positions in relative polar coordinate system rather than the Cartesian one. Despite being pre-trained on a dataset more than six times smaller than the widely used IIT-CDIP corpus, DocPolarBERT achieves state-of-the-art results. These results demonstrate that a carefully designed attention mechanism can compensate for reduced pre-training data, offering an efficient and effective alternative for document understanding.
♻ ☆ DRAGON: Dynamic RAG Benchmark On News
Retrieval-Augmented Generation (RAG) is a widely adopted approach for improving the factuality of large language models (LLMs) by incorporating external knowledge at inference time. Although there exist multiple RAG benchmarks for English, evaluation resources for other languages, including Russian, remain scarce and static, failing to capture the dynamic nature of real-world deployments. In this work, we present DRAGON (Dynamic RAG Benchmark On News), the first dynamic benchmark for evaluating RAG systems in Russian on a changing news corpora. DRAGON is built upon a regularly updated corpus of Russian news and public documents and supports comprehensive evaluation of both the retriever and generator components. Question generation is performed automatically with the use of Knowledge Graph constructed from the corpus and enables the extraction of four core question types aligned with distinct subgraph patterns. We release a complete evaluation framework comprising the pipeline for automatic question generation, evaluation scripts, which are potentially reusable for other languages and multilingual settings, and benchmark data. We also launch a public leaderboard to encourage community participation and comparison.
♻ ☆ Understanding the Dark Side of LLMs' Intrinsic Self-Correction
Intrinsic self-correction was proposed to improve LLMs' responses via feedback prompts solely based on their inherent capability. However, recent works show that LLMs' intrinsic self-correction fails without oracle labels as feedback prompts. In this paper, we aim to interpret LLMs' intrinsic self-correction for different tasks, especially for those failure cases. By including one simple task and three complex tasks with state-of-the-art (SOTA) LLMs like ChatGPT families (o1, 4o, 3.5-turbo) and Llama families (2-7B, 3-8B, and 3.1-8B), we design three interpretation methods to reveal the dark side of LLMs' intrinsic self-correction. We identify intrinsic self-correction can (1) cause LLMs to waver both intermedia and final answers and lead to prompt bias on simple factual questions; (2) introduce human-like cognitive bias on complex tasks. In light of our findings, we also provide two simple yet effective strategies for alleviation: question repeating and supervised fine-tuning with a few samples. We open-source our work at https://x-isc.info/.
♻ ☆ ReVISE: Learning to Refine at Test-Time via Intrinsic Self-Verification ICML 2025
Self-awareness, i.e., the ability to assess and correct one's own generation, is a fundamental aspect of human intelligence, making its replication in large language models (LLMs) an important yet challenging task. Previous works tackle this by employing extensive reinforcement learning or rather relying on large external verifiers. In this work, we propose Refine via Intrinsic Self-Verification (ReVISE), an efficient and effective framework that enables LLMs to self-correct their outputs through self-verification. The core idea of ReVISE is to enable LLMs to verify their reasoning processes and continually rethink reasoning trajectories based on its verification. We introduce a structured curriculum based upon online preference learning to implement this efficiently. Specifically, as ReVISE involves two challenging tasks (i.e., self-verification and reasoning correction), we tackle each task sequentially using curriculum learning, collecting both failed and successful reasoning paths to construct preference pairs for efficient training. During inference, our approach enjoys natural test-time scaling by integrating self-verification and correction capabilities, further enhanced by our proposed confidence-aware decoding mechanism. Our experiments on various reasoning tasks demonstrate that ReVISE achieves efficient self-correction and significantly improves reasoning performance.
comment: Published as conference proceeding for ICML 2025. First two authors contributed equally
♻ ☆ REST: Stress Testing Large Reasoning Models by Asking Multiple Problems at Once
Recent Large Reasoning Models (LRMs) have achieved remarkable progress on task-specific benchmarks, yet their evaluation methods remain constrained by isolated problem-solving paradigms. Existing benchmarks predominantly assess single-question reasoning through sequential testing, resulting critical limitations: (1) vulnerability to data contamination and less challenging (e.g., DeepSeek-R1 achieves 97.0% on MATH500), forcing costly creation of new questions with large human efforts, (2) failure to evaluate models under multi-context pressure, a key requirement for real-world deployment. To bridge this gap, we present REST (Reasoning Evaluation through Simultaneous Testing), a stress-testing framework that exposes LRMs to multiple problems simultaneously. Beyond basic reasoning, REST evaluates several under-tested capabilities: contextual priority allocation, cross-problem interference resistance, and dynamic cognitive load management. Our evaluation reveals several striking findings: Even state-of-the-art (SOTA) models like DeepSeek-R1 exhibit substantial performance degradation under stress testing. Crucially, REST demonstrates stronger discriminative power than existing benchmarks, revealing pronounced performance differences among models that exhibit similar, near-ceiling performance under single-question evaluations. Some key insights emerge from our analysis: (1) the "overthinking trap" is a critical factor contributing to the performance degradation; (2) the models trained with "long2short" technique preserve more accuracy of their single-problem performance under REST, outperforming standard-trained counterparts. These results establish REST as a cost-efficient, future-proof evaluation paradigm that better reflects real-world reasoning demands while reducing reliance on continuous human annotation. Code and results are available at https://opendatalab.github.io/REST.
comment: REST (Reasoning Evaluation through Simultaneous Testing), a stress-testing framework that concurrently exposes LRMs to multiple problems simultaneously
♻ ☆ Bring Reason to Vision: Understanding Perception and Reasoning through Model Merging ICML 2025
Vision-Language Models (VLMs) combine visual perception with the general capabilities, such as reasoning, of Large Language Models (LLMs). However, the mechanisms by which these two abilities can be combined and contribute remain poorly understood. In this work, we explore to compose perception and reasoning through model merging that connects parameters of different models. Unlike previous works that often focus on merging models of the same kind, we propose merging models across modalities, enabling the incorporation of the reasoning capabilities of LLMs into VLMs. Through extensive experiments, we demonstrate that model merging offers a successful pathway to transfer reasoning abilities from LLMs to VLMs in a training-free manner. Moreover, we utilize the merged models to understand the internal mechanism of perception and reasoning and how merging affects it. We find that perception capabilities are predominantly encoded in the early layers of the model, whereas reasoning is largely facilitated by the middle-to-late layers. After merging, we observe that all layers begin to contribute to reasoning, whereas the distribution of perception abilities across layers remains largely unchanged. These observations shed light on the potential of model merging as a tool for multimodal integration and interpretation.
comment: ICML 2025. Camera-ready version updated. Our code is publicly available at https://github.com/shiqichen17/VLM_Merging
♻ ☆ Leveraging Large Language Models for Multi-Class and Multi-Label Detection of Drug Use and Overdose Symptoms on Social Media
Drug overdose remains a critical global health issue, often driven by misuse of opioids, painkillers, and psychiatric medications. Traditional research methods face limitations, whereas social media offers real-time insights into self-reported substance use and overdose symptoms. This study proposes an AI-driven NLP framework trained on annotated social media data to detect commonly used drugs and associated overdose symptoms. Using a hybrid annotation strategy with LLMs and human annotators, we applied traditional ML models, neural networks, and advanced transformer-based models. Our framework achieved 98% accuracy in multi-class and 97% in multi-label classification, outperforming baseline models by up to 8%. These findings highlight the potential of AI for supporting public health surveillance and personalized intervention strategies.
♻ ☆ Online Intrinsic Rewards for Decision Making Agents from Large Language Model Feedback
Automatically synthesizing dense rewards from natural language descriptions is a promising paradigm in reinforcement learning (RL), with applications to sparse reward problems, open-ended exploration, and hierarchical skill design. Recent works have made promising steps by exploiting the prior knowledge of large language models (LLMs). However, these approaches suffer from important limitations: they are either not scalable to problems requiring billions of environment samples, due to requiring LLM annotations for each observation, or they require a diverse offline dataset, which may not exist or be impossible to collect. In this work, we address these limitations through a combination of algorithmic and systems-level contributions. We propose ONI, a distributed architecture that simultaneously learns an RL policy and an intrinsic reward function using LLM feedback. Our approach annotates the agent's collected experience via an asynchronous LLM server, which is then distilled into an intrinsic reward model. We explore a range of algorithmic choices for reward modeling with varying complexity, including hashing, classification, and ranking models. Our approach achieves state-of-the-art performance across a range of challenging tasks from the NetHack Learning Environment, while removing the need for large offline datasets required by prior work. We make our code available at https://github.com/facebookresearch/oni .
♻ ☆ BMDetect: A Multimodal Deep Learning Framework for Comprehensive Biomedical Misconduct Detection
Academic misconduct detection in biomedical research remains challenging due to algorithmic narrowness in existing methods and fragmented analytical pipelines. We present BMDetect, a multimodal deep learning framework that integrates journal metadata (SJR, institutional data), semantic embeddings (PubMedBERT), and GPT-4o-mined textual attributes (methodological statistics, data anomalies) for holistic manuscript evaluation. Key innovations include: (1) multimodal fusion of domain-specific features to reduce detection bias; (2) quantitative evaluation of feature importance, identifying journal authority metrics (e.g., SJR-index) and textual anomalies (e.g., statistical outliers) as dominant predictors; and (3) the BioMCD dataset, a large-scale benchmark with 13,160 retracted articles and 53,411 controls. BMDetect achieves 74.33% AUC, outperforming single-modality baselines by 8.6%, and demonstrates transferability across biomedical subfields. This work advances scalable, interpretable tools for safeguarding research integrity.
♻ ☆ SEALGuard: Safeguarding the Multilingual Conversations in Southeast Asian Languages for LLM Software Systems
Safety alignment is critical for LLM-powered systems. While recent LLM-powered guardrail approaches such as LlamaGuard achieve high detection accuracy of unsafe inputs written in English (e.g., ``How to create a bomb?''), they struggle with multilingual unsafe inputs. This limitation leaves LLM systems vulnerable to unsafe and jailbreak prompts written in low-resource languages such as those in Southeast Asia. This paper introduces SEALGuard, a multilingual guardrail designed to improve the safety alignment across diverse languages. It aims to address the multilingual safety alignment gap of existing guardrails and ensure effective filtering of unsafe and jailbreak prompts in LLM-powered systems. We adapt a general-purpose multilingual language model into a multilingual guardrail using low-rank adaptation (LoRA). We construct SEALSBench, a large-scale multilingual safety alignment dataset containing over 260,000 prompts in ten languages, including safe, unsafe, and jailbreak cases. We evaluate SEALGuard against state-of-the-art guardrails such as LlamaGuard on this benchmark. Our findings show that multilingual unsafe and jailbreak prompts substantially degrade the performance of the state-of-the-art LlamaGuard, which experiences a drop in Defense Success Rate (DSR) by 9% and 18%, respectively, compared to its performance on English-only prompts. In contrast, SEALGuard outperforms existing guardrails in detecting multilingual unsafe and jailbreak prompts, improving DSR by 48% over LlamaGuard and achieving the best DSR, precision, and F1-score. Our ablation study further reveals the contributions of adaptation strategies and model size to the overall performance of SEALGuard. SEALGuard advances the safety alignment of LLM systems by introducing an effective multilingual guardrail.
comment: Under Review at Information and Software Technology
♻ ☆ Is Training Data Quality or Quantity More Impactful to Small Language Model Performance?
This study investigates the relative impact of training data quality versus quantity on the performance of small language models (SLMs), utilizing the TinyStories dataset for empirical analysis. Analysis of dataset variations with respect to size (25% and 50% of the original size) and duplication (controlled rates of 25%, 50%, 75%, and 100%) were performed. Model performance was evaluated based on the validation loss, accuracy, and perplexity metrics. Results indicate training data quality plays a more significant role in the overall performance of SLMs, especially given scale of this experiment. Minimal duplication positively impacted model accuracy (+0.87% increase in accuracy at 25% duplication) without significantly increasing perplexity (+0.52% increase going from 0% to 25% duplication) but excessive duplication led to pronounced performance degradation (-40% drop in accuracy at 100% duplication). The implications of this exploration extend beyond just model performance; training large-scale models imposes significant financial and computational burdens, which can be prohibitive for organizations, individuals, and the public at large, especially in developing countries. Additionally, the energy consumption associated with large-scale training raises environmental concerns. Understanding the relative importance of data quality versus quantity could democratize AI technology, making advanced models more accessible and sustainable for all.
comment: 14 pages, 5 tables, 4 figures | Accepted at International Conference on Neural Computing for Advanced Applications 2025, Conference info: https://aaci.org.hk/ncaa2025
♻ ☆ Prompt4Trust: A Reinforcement Learning Prompt Augmentation Framework for Clinically-Aligned Confidence Calibration in Multimodal Large Language Models ICCV 2025
Multimodal large language models (MLLMs) hold considerable promise for applications in healthcare. However, their deployment in safety-critical settings is hindered by two key limitations: (i) sensitivity to prompt design, and (ii) a tendency to generate incorrect responses with high confidence. As clinicians may rely on a model's stated confidence to gauge the reliability of its predictions, it is especially important that when a model expresses high confidence, it is also highly accurate. We introduce Prompt4Trust, the first reinforcement learning (RL) framework for prompt augmentation targeting confidence calibration in MLLMs. A lightweight LLM is trained to produce context-aware auxiliary prompts that guide a downstream task MLLM to generate responses in which the expressed confidence more accurately reflects predictive accuracy. Unlike conventional calibration techniques, Prompt4Trust specifically prioritizes aspects of calibration most critical for safe and trustworthy clinical decision-making. Beyond improvements driven by this clinically motivated calibration objective, our proposed method also improves task accuracy, achieving state-of-the-art medical visual question answering (VQA) performance on the PMC-VQA benchmark, which is composed of multiple-choice questions spanning diverse medical imaging modalities. Moreover, our framework trained with a small downstream task MLLM showed promising zero-shot generalization to larger MLLMs in our experiments, suggesting the potential for scalable calibration without the associated computational costs. This work demonstrates the potential of automated yet human-aligned prompt engineering for improving the the trustworthiness of MLLMs in safety critical settings. Our codebase can be found at https://github.com/xingbpshen/prompt4trust.
comment: Accepted to ICCV 2025 Workshop CVAMD
♻ ☆ Representation Bending for Large Language Model Safety ACL 2025
Large Language Models (LLMs) have emerged as powerful tools, but their inherent safety risks - ranging from harmful content generation to broader societal harms - pose significant challenges. These risks can be amplified by the recent adversarial attacks, fine-tuning vulnerabilities, and the increasing deployment of LLMs in high-stakes environments. Existing safety-enhancing techniques, such as fine-tuning with human feedback or adversarial training, are still vulnerable as they address specific threats and often fail to generalize across unseen attacks, or require manual system-level defenses. This paper introduces RepBend, a novel approach that fundamentally disrupts the representations underlying harmful behaviors in LLMs, offering a scalable solution to enhance (potentially inherent) safety. RepBend brings the idea of activation steering - simple vector arithmetic for steering model's behavior during inference - to loss-based fine-tuning. Through extensive evaluation, RepBend achieves state-of-the-art performance, outperforming prior methods such as Circuit Breaker, RMU, and NPO, with up to 95% reduction in attack success rates across diverse jailbreak benchmarks, all with negligible reduction in model usability and general capabilities.
comment: Accepted to ACL 2025 (main)
♻ ☆ The GPT Surprise: Offering Large Language Model Chat in a Massive Coding Class Reduced Engagement but Increased Adopters Exam Performances
Large language models (LLMs) are quickly being adopted in a wide range of learning experiences, especially via ubiquitous and broadly accessible chat interfaces like ChatGPT and Copilot. This type of interface is readily available to students and teachers around the world, yet relatively little research has been done to assess the impact of such generic tools on student learning. Coding education is an interesting test case, both because LLMs have strong performance on coding tasks, and because LLM-powered support tools are rapidly becoming part of the workflow of professional software engineers. To help understand the impact of generic LLM use on coding education, we conducted a large-scale randomized control trial with 5,831 students from 146 countries in an online coding class in which we provided some students with access to a chat interface with GPT-4. We estimate positive benefits on exam performance for adopters, the students who used the tool, but over all students, the advertisement of GPT-4 led to a significant average decrease in exam participation. We observe similar decreases in other forms of course engagement. However, this decrease is modulated by the student's country of origin. Offering access to LLMs to students from low human development index countries increased their exam participation rate on average. Our results suggest there may be promising benefits to using LLMs in an introductory coding class, but also potential harms for engagement, which makes their longer term impact on student success unclear. Our work highlights the need for additional investigations to help understand the potential impact of future adoption and integration of LLMs into classrooms.
comment: 32 pages. Published at L@S 2025
♻ ☆ Fine-grained Stateful Knowledge Exploration: Effective and Efficient Graph Retrieval with Large Language Models
Large Language Models (LLMs) have shown impressive capabilities, yet updating their knowledge remains a significant challenge, often leading to outdated or inaccurate responses. A proposed solution is the integration of external knowledge bases, such as knowledge graphs, with LLMs. Most existing methods use a paradigm that treats the whole question as the objective, with relevant knowledge being incrementally retrieved from the knowledge graph. However, this paradigm often leads to a granularity mismatch between the target question and the retrieved entities and relations. As a result, the information in the question cannot precisely correspond to the retrieved knowledge. This may cause redundant exploration or omission of vital knowledge, thereby leading to enhanced computational consumption and reduced retrieval accuracy. To address the limitations of coarse-grained knowledge exploration, we propose FiSKE, a novel paradigm for Fine-grained Stateful Knowledge Exploration. FiSKE first decomposes questions into fine-grained clues, then employs an adaptive mapping strategy during knowledge exploration process to resolve ambiguity in clue-to-graph mappings. This strategy dynamically infers contextual correspondences while maintaining a stateful record of the mappings. A clue-driven termination mechanism ensures rigorous augmentation--leveraging fully mapped paths for LLMs while reverting to chain-of-thought reasoning when necessary. Our approach balances precision and efficiency. Experiments on multiple datasets revealed that our paradigm surpasses current advanced methods in knowledge retrieval while significantly reducing the average number of LLM invocations.
♻ ☆ Multimodal Sentiment Analysis on CMU-MOSEI Dataset using Transformer-based Models
This project performs multimodal sentiment analysis using the CMU-MOSEI dataset, using transformer-based models with early fusion to integrate text, audio, and visual modalities. We employ BERT-based encoders for each modality, extracting embeddings that are concatenated before classification. The model achieves strong performance, with 97.87% 7-class accuracy and a 0.9682 F1-score on the test set, demonstrating the effectiveness of early fusion in capturing cross-modal interactions. The training utilized Adam optimization (lr=1e-4), dropout (0.3), and early stopping to ensure generalization and robustness. Results highlight the superiority of transformer architectures in modeling multimodal sentiment, with a low MAE (0.1060) indicating precise sentiment intensity prediction. Future work may compare fusion strategies or enhance interpretability. This approach utilizes multimodal learning by effectively combining linguistic, acoustic, and visual cues for sentiment analysis.
comment: 6 pages, 2 figures
♻ ☆ ZebraLogic: On the Scaling Limits of LLMs for Logical Reasoning ICML 2025
We investigate the logical reasoning capabilities of large language models (LLMs) and their scalability in complex non-monotonic reasoning. To this end, we introduce ZebraLogic, a comprehensive evaluation framework for assessing LLM reasoning performance on logic grid puzzles derived from constraint satisfaction problems (CSPs). ZebraLogic enables the generation of puzzles with controllable and quantifiable complexity, facilitating a systematic study of the scaling limits of models such as Llama, o1 models, and DeepSeek-R1. By encompassing a broad range of search space complexities and diverse logical constraints, ZebraLogic provides a structured environment to evaluate reasoning under increasing difficulty. Our results reveal a significant decline in accuracy as problem complexity grows -- a phenomenon we term the curse of complexity. This limitation persists even with larger models and increased inference-time computation, suggesting inherent constraints in current LLM reasoning capabilities. Additionally, we explore strategies to enhance logical reasoning, including Best-of-N sampling, backtracking mechanisms, and self-verification prompts. Our findings offer critical insights into the scalability of LLM reasoning, highlight fundamental limitations, and outline potential directions for improvement.
comment: Accepted to ICML 2025
♻ ☆ GenARM: Reward Guided Generation with Autoregressive Reward Model for Test-time Alignment ICLR 2025
Large Language Models (LLMs) exhibit impressive capabilities but require careful alignment with human preferences. Traditional training-time methods finetune LLMs using human preference datasets but incur significant training costs and require repeated training to handle diverse user preferences. Test-time alignment methods address this by using reward models (RMs) to guide frozen LLMs without retraining. However, existing test-time approaches rely on trajectory-level RMs which are designed to evaluate complete responses, making them unsuitable for autoregressive text generation that requires computing next-token rewards from partial responses. To address this, we introduce GenARM, a test-time alignment approach that leverages the Autoregressive Reward Model--a novel reward parametrization designed to predict next-token rewards for efficient and effective autoregressive generation. Theoretically, we demonstrate that this parametrization can provably guide frozen LLMs toward any distribution achievable by traditional RMs within the KL-regularized reinforcement learning framework. Experimental results show that GenARM significantly outperforms prior test-time alignment baselines and matches the performance of training-time methods. Additionally, GenARM enables efficient weak-to-strong guidance, aligning larger LLMs with smaller RMs without the high costs of training larger models. Furthermore, GenARM supports multi-objective alignment, allowing real-time trade-offs between preference dimensions and catering to diverse user preferences without retraining. Our project page is available at: https://genarm.github.io.
comment: Published at the Thirteenth International Conference on Learning Representations (ICLR 2025)
♻ ☆ Jan-nano Technical Report
Most language models face a fundamental tradeoff where powerful capabilities require substantial computational resources. We shatter this constraint with Jan-nano, a 4B parameter language model that redefines efficiency through radical specialization: instead of trying to know everything, it masters the art of finding anything instantly. Fine-tuned from Qwen3-4B using our novel multi-stage Reinforcement Learning with Verifiable Rewards (RLVR) system that completely eliminates reliance on next token prediction training (SFT), Jan-nano achieves 83.2% on SimpleQA benchmark with MCP integration while running on consumer hardware. With 128K context length, Jan-nano proves that intelligence isn't about scale, it's about strategy.
♻ ☆ A quantum semantic framework for natural language processing
Semantic degeneracy represents a fundamental property of natural language that extends beyond simple polysemy to encompass the combinatorial explosion of potential interpretations that emerges as semantic expressions increase in complexity. In this work, we argue this property imposes fundamental limitations on Large Language Models (LLMs) and other modern NLP systems, precisely because they operate within natural language itself. Using Kolmogorov complexity, we demonstrate that as an expression's complexity grows, the amount of contextual information required to reliably resolve its ambiguity explodes combinatorially. The computational intractability of recovering a single intended meaning for complex or ambiguous text therefore suggests that the classical view that linguistic forms possess intrinsic meaning in and of themselves is conceptually inadequate. We argue instead that meaning is dynamically actualized through an observer-dependent interpretive act, a process whose non-deterministic nature is most appropriately described by a non-classical, quantum-like logic. To test this hypothesis, we conducted a semantic Bell inequality test using diverse LLM agents. Our experiments yielded average CHSH expectation values from 1.2 to 2.8, with several runs producing values (e.g., 2.3-2.4) in significant violation of the classical boundary ($|S|\leq2$), demonstrating that linguistic interpretation under ambiguity can exhibit non-classical contextuality, consistent with results from human cognition experiments. These results inherently imply that classical frequentist-based analytical approaches for natural language are necessarily lossy. Instead, we propose that Bayesian-style repeated sampling approaches can provide more practically useful and appropriate characterizations of linguistic meaning in context.
comment: 12 pages, 2 figures, accepted submission to Quantum AI and NLP 2025
♻ ☆ How Well Can Knowledge Edit Methods Edit Perplexing Knowledge?
Large language models (LLMs) have demonstrated remarkable capabilities, but updating their knowledge post-training remains a critical challenge. While recent model editing techniques like Rank-One Model Editing (ROME) show promise, their effectiveness may vary based on the nature of the knowledge being edited. We introduce the concept of ``perplexingness'': the degree to which new knowledge conflicts with an LLM's learned conceptual hierarchies and categorical relationships. For instance, editing ``British Shorthair is a kind of cat'' to ``British Shorthair is a kind of dog'' represents a low-perplexingness edit within the same taxonomic level, while editing ``A cat is a kind of animal'' to ``A cat is a kind of plant'' represents a high-perplexingness edit that violates fundamental categorical boundaries. To systematically investigate this phenomenon, we introduce HierarchyData, a carefully curated dataset of 99 hyponym-hypernym pairs across diverse categories. Through controlled experiments across three models and four editing methods, we demonstrate a strong negative correlation between the perplexingness of new knowledge and the effectiveness of knowledge editing. Our analysis reveals that edits involving more abstract concepts (hypernyms) generally exhibit higher perplexingness and are more resistant to modification than their specific counterparts (hyponyms). These findings highlight a fundamental challenge in LLM knowledge editing: the more a new fact contradicts an LLM's learned conceptual hierarchies, the harder it becomes to reliably encode that knowledge.
comment: A previous version of this document contained a hidden prompt entered by Z Zhu without knowledge of -- or consent by -- his co-authors. This version does not contain the prompt
♻ ☆ Understanding Language Model Circuits through Knowledge Editing
Recent advances in language model interpretability have identified circuits, critical subnetworks that replicate model behaviors, yet how knowledge is structured within these crucial subnetworks remains opaque. To gain an understanding toward the knowledge in the circuits, we conduct systematic knowledge editing experiments on the circuits of the GPT-2 language model. Our analysis reveals intriguing patterns in how circuits respond to editing attempts, the extent of knowledge distribution across network components, and the architectural composition of knowledge-bearing circuits. These findings offer insights into the complex relationship between model circuits and knowledge representation, deepening the understanding of how information is organized within language models. Our findings offer novel insights into the ``meanings'' of the circuits, and introduce directions for further interpretability and safety research of language models.
comment: A previous version of this document contained a hidden prompt entered by Z Zhu without knowledge of -- or consent by -- his co-authors. This version does not contain the prompt
♻ ☆ AKReF: An argumentative knowledge representation framework for structured argumentation
This paper presents a framework to convert argumentative texts into argument knowledge graphs (AKG). The proposed argumentative knowledge representation framework (AKReF) extends the theoretical foundation and enables the AKG to provide a graphical view of the argumentative structure that is easier to understand. Starting with basic annotations of argumentative components (ACs) and argumentative relations (ARs), we enrich the information by constructing a knowledge base (KB) graph with metadata attributes for nodes. Next, we apply modus ponens on premises and inference rules from the KB to form arguments. From these arguments, we create an AKG. The nodes and edges of the AKG have attributes capturing key argumentative features such as the type of premise (e.g., axiom, ordinary premise, assumption), the type of inference rule (e.g., strict, defeasible), preference order over defeasible rules, markers (e.g., "therefore", "however"), and the type of attack (e.g., undercut, rebuttal, undermining). We identify inference rules by locating a specific set of markers, called inference markers (IM). This, in turn, makes it possible to identify undercut attacks previously undetectable in existing datasets. AKG prepares the ground for reasoning tasks, including checking the coherence of arguments and identifying opportunities for revision. For this, it is essential to find indirect relations, many of which are implicit. Our proposed AKG format, with annotated inference rules and modus ponens, helps reasoning models learn the implicit, indirect relations that require inference over arguments and their interconnections. We use an essay from the AAEC dataset to illustrate the framework. We further show its application in complex analyses such as extracting a conflict-free set and a maximal set of admissible arguments.
comment: 20 pages, 7 figures, 2 tables
♻ ☆ Flexible and Efficient Grammar-Constrained Decoding
Large Language Models (LLMs) are often asked to generate structured outputs that obey precise syntactic rules, such as code snippets or formatted data. Grammar-constrained decoding (GCD) can guarantee that LLM outputs matches such rules by masking out tokens that will provably lead to outputs that do not belong to a specified context-free grammar (CFG). To guarantee soundness, GCD algorithms have to compute how a given LLM subword tokenizer can align with the tokens used by a given context-free grammar and compute token masks based on this information. Doing so efficiently is challenging and existing GCD algorithms require tens of minutes to preprocess common grammars. We present a new GCD algorithm together with an implementation that offers 17.71x faster offline preprocessing than existing approaches while preserving state-of-the-art efficiency in online mask computation.
♻ ☆ Multi-domain Multilingual Sentiment Analysis in Industry: Predicting Aspect-based Opinion Quadruples
This paper explores the design of an aspect-based sentiment analysis system using large language models (LLMs) for real-world use. We focus on quadruple opinion extraction -- identifying aspect categories, sentiment polarity, targets, and opinion expressions from text data across different domains and languages. We investigate whether a single fine-tuned model can effectively handle multiple domain-specific taxonomies simultaneously. We demonstrate that a combined multi-domain model achieves performance comparable to specialized single-domain models while reducing operational complexity. We also share lessons learned for handling non-extractive predictions and evaluating various failure modes when developing LLM-based systems for structured prediction tasks.
♻ ☆ Rolling the DICE on Idiomaticity: How LLMs Fail to Grasp Context ACL 2025
Human processing of idioms relies on understanding the contextual sentences in which idioms occur, as well as language-intrinsic features such as frequency and speaker-intrinsic factors like familiarity. While LLMs have shown high performance on idiomaticity detection tasks, this success may be attributed to reasoning shortcuts in existing datasets. To this end, we construct a novel, controlled contrastive dataset designed to test whether LLMs can effectively use context to disambiguate idiomatic meaning. Additionally, we explore how collocational frequency and sentence probability influence model performance. Our findings reveal that LLMs often fail to resolve idiomaticity when it is required to attend to the surrounding context, and that models perform better on sentences that have higher likelihood. The collocational frequency of expressions also impacts performance. We make our code and dataset publicly available.
comment: ACL 2025
♻ ☆ Efficient and Direct Duplex Modeling for Speech-to-Speech Language Model
Spoken dialogue is an intuitive form of human-computer interaction, yet current speech language models often remain constrained to turn-based exchanges, lacking real-time adaptability such as user barge-in. We propose a novel duplex speech to speech (S2S) architecture featuring continuous user inputs and codec agent outputs with channel fusion that directly models simultaneous user and agent streams. Using a pretrained streaming encoder for user input enables the first duplex S2S model without requiring speech pretrain. Separate architectures for agent and user modeling facilitate codec fine-tuning for better agent voices and halve the bitrate (0.6 kbps) compared to previous works. Experimental results show that the proposed model outperforms previous duplex models in reasoning, turn-taking, and barge-in abilities. The model requires significantly less speech data, as speech pretrain is skipped, which markedly simplifies the process of building a duplex S2S model from any LLMs. Finally, it is the first openly available duplex S2S model with training and inference code to foster reproducibility.
comment: Accepted to Interspeech 2025
Computer Vision and Pattern Recognition 15
☆ Towards Depth Foundation Model: Recent Trends in Vision-Based Depth Estimation
Depth estimation is a fundamental task in 3D computer vision, crucial for applications such as 3D reconstruction, free-viewpoint rendering, robotics, autonomous driving, and AR/VR technologies. Traditional methods relying on hardware sensors like LiDAR are often limited by high costs, low resolution, and environmental sensitivity, limiting their applicability in real-world scenarios. Recent advances in vision-based methods offer a promising alternative, yet they face challenges in generalization and stability due to either the low-capacity model architectures or the reliance on domain-specific and small-scale datasets. The emergence of scaling laws and foundation models in other domains has inspired the development of "depth foundation models": deep neural networks trained on large datasets with strong zero-shot generalization capabilities. This paper surveys the evolution of deep learning architectures and paradigms for depth estimation across the monocular, stereo, multi-view, and monocular video settings. We explore the potential of these models to address existing challenges and provide a comprehensive overview of large-scale datasets that can facilitate their development. By identifying key architectures and training strategies, we aim to highlight the path towards robust depth foundation models, offering insights into their future research and applications.
☆ Streaming 4D Visual Geometry Transformer
Perceiving and reconstructing 4D spatial-temporal geometry from videos is a fundamental yet challenging computer vision task. To facilitate interactive and real-time applications, we propose a streaming 4D visual geometry transformer that shares a similar philosophy with autoregressive large language models. We explore a simple and efficient design and employ a causal transformer architecture to process the input sequence in an online manner. We use temporal causal attention and cache the historical keys and values as implicit memory to enable efficient streaming long-term 4D reconstruction. This design can handle real-time 4D reconstruction by incrementally integrating historical information while maintaining high-quality spatial consistency. For efficient training, we propose to distill knowledge from the dense bidirectional visual geometry grounded transformer (VGGT) to our causal model. For inference, our model supports the migration of optimized efficient attention operator (e.g., FlashAttention) from the field of large language models. Extensive experiments on various 4D geometry perception benchmarks demonstrate that our model increases the inference speed in online scenarios while maintaining competitive performance, paving the way for scalable and interactive 4D vision systems. Code is available at: https://github.com/wzzheng/StreamVGGT.
comment: Code is available at: https://github.com/wzzheng/StreamVGGT
☆ CharaConsist: Fine-Grained Consistent Character Generation ICCV 2025
In text-to-image generation, producing a series of consistent contents that preserve the same identity is highly valuable for real-world applications. Although a few works have explored training-free methods to enhance the consistency of generated subjects, we observe that they suffer from the following problems. First, they fail to maintain consistent background details, which limits their applicability. Furthermore, when the foreground character undergoes large motion variations, inconsistencies in identity and clothing details become evident. To address these problems, we propose CharaConsist, which employs point-tracking attention and adaptive token merge along with decoupled control of the foreground and background. CharaConsist enables fine-grained consistency for both foreground and background, supporting the generation of one character in continuous shots within a fixed scene or in discrete shots across different scenes. Moreover, CharaConsist is the first consistent generation method tailored for text-to-image DiT model. Its ability to maintain fine-grained consistency, combined with the larger capacity of latest base model, enables it to produce high-quality visual outputs, broadening its applicability to a wider range of real-world scenarios. The source code has been released at https://github.com/Murray-Wang/CharaConsist
comment: ICCV 2025 accepted paper, project page: https://murray-wang.github.io/CharaConsist/
☆ CATVis: Context-Aware Thought Visualization MICCAI 2025
EEG-based brain-computer interfaces (BCIs) have shown promise in various applications, such as motor imagery and cognitive state monitoring. However, decoding visual representations from EEG signals remains a significant challenge due to their complex and noisy nature. We thus propose a novel 5-stage framework for decoding visual representations from EEG signals: (1) an EEG encoder for concept classification, (2) cross-modal alignment of EEG and text embeddings in CLIP feature space, (3) caption refinement via re-ranking, (4) weighted interpolation of concept and caption embeddings for richer semantics, and (5) image generation using a pre-trained Stable Diffusion model. We enable context-aware EEG-to-image generation through cross-modal alignment and re-ranking. Experimental results demonstrate that our method generates high-quality images aligned with visual stimuli, outperforming SOTA approaches by 13.43% in Classification Accuracy, 15.21% in Generation Accuracy and reducing Fr\'echet Inception Distance by 36.61%, indicating superior semantic alignment and image quality.
comment: Accepted at MICCAI 2025. This is the submitted version prior to peer review. The final Version of Record will appear in the MICCAI 2025 proceedings (Springer LNCS)
☆ COLIBRI Fuzzy Model: Color Linguistic-Based Representation and Interpretation
Colors are omnipresent in today's world and play a vital role in how humans perceive and interact with their surroundings. However, it is challenging for computers to imitate human color perception. This paper introduces the Human Perception-Based Fuzzy Color Model, COLIBRI (Color Linguistic-Based Representation and Interpretation), designed to bridge the gap between computational color representations and human visual perception. The proposed model uses fuzzy sets and logic to create a framework for color categorization. Using a three-phase experimental approach, the study first identifies distinguishable color stimuli for hue, saturation, and intensity through preliminary experiments, followed by a large-scale human categorization survey involving more than 1000 human subjects. The resulting data are used to extract fuzzy partitions and generate membership functions that reflect real-world perceptual uncertainty. The model incorporates a mechanism for adaptation that allows refinement based on feedback and contextual changes. Comparative evaluations demonstrate the model's alignment with human perception compared to traditional color models, such as RGB, HSV, and LAB. To the best of our knowledge, no previous research has documented the construction of a model for color attribute specification based on a sample of this size or a comparable sample of the human population (n = 2496). Our findings are significant for fields such as design, artificial intelligence, marketing, and human-computer interaction, where perceptually relevant color representation is critical.
comment: submitted to IEEE for consideration
☆ C-FBI: A Combinatorial method using Convolutions for Circle Fitting in Blurry Images
This paper addresses the fundamental computer vision challenge of robust circle detection and fitting in degraded imaging conditions. We present Combinatorial Convolution-based Circle Fitting for Blurry Images (3C-FBI), an algorithm that bridges the gap between circle detection and precise parametric fitting by combining (1) efficient combinatorial edge pixel (edgel) sampling and (2) convolution-based density estimation in parameter space. We evaluate 3C-FBI across three experimental frameworks: (1) real-world medical data from Parkinson's disease assessments (144 frames from 36 videos), (2) controlled synthetic data following established circle-fitting benchmarks, and (3) systematic analysis across varying spatial resolutions and outlier contamination levels. Results show that 3C-FBI achieves state-of-the-art accuracy (Jaccard index 0.896) while maintaining real-time performance (40.3 fps), significantly outperforming classical methods like RCD (6.8 fps) on a standard CPU (i7-10875H). It maintains near-perfect accuracy (Jaccard almost 1.0) at high resolutions (480x480) and reliable performance (Jaccard higher than 0.95) down to 160x160 with up to 20% outliers. In extensive synthetic testing, 3C-FBI achieves a mean Jaccard Index of 0.989 across contamination levels, comparable to modern methods like Qi et al. (2024, 0.991), and surpassing RHT (0.964). This combination of accuracy, speed, and robustness makes 3C-FBI ideal for medical imaging, robotics, and industrial inspection under challenging conditions.
comment: 22 pages, 16 figures
☆ HUG-VAS: A Hierarchical NURBS-Based Generative Model for Aortic Geometry Synthesis and Controllable Editing
Accurate characterization of vascular geometry is essential for cardiovascular diagnosis and treatment planning. Traditional statistical shape modeling (SSM) methods rely on linear assumptions, limiting their expressivity and scalability to complex topologies such as multi-branch vascular structures. We introduce HUG-VAS, a Hierarchical NURBS Generative model for Vascular geometry Synthesis, which integrates NURBS surface parameterization with diffusion-based generative modeling to synthesize realistic, fine-grained aortic geometries. Trained with 21 patient-specific samples, HUG-VAS generates anatomically faithful aortas with supra-aortic branches, yielding biomarker distributions that closely match those of the original dataset. HUG-VAS adopts a hierarchical architecture comprising a denoising diffusion model that generates centerlines and a guided diffusion model that synthesizes radial profiles conditioned on those centerlines, thereby capturing two layers of anatomical variability. Critically, the framework supports zero-shot conditional generation from image-derived priors, enabling practical applications such as interactive semi-automatic segmentation, robust reconstruction under degraded imaging conditions, and implantable device optimization. To our knowledge, HUG-VAS is the first SSM framework to bridge image-derived priors with generative shape modeling via a unified integration of NURBS parameterization and hierarchical diffusion processes.
comment: 59 pages, 9 figures
☆ Elevating 3D Models: High-Quality Texture and Geometry Refinement from a Low-Quality Model SIGGRAPH 2025
High-quality 3D assets are essential for various applications in computer graphics and 3D vision but remain scarce due to significant acquisition costs. To address this shortage, we introduce Elevate3D, a novel framework that transforms readily accessible low-quality 3D assets into higher quality. At the core of Elevate3D is HFS-SDEdit, a specialized texture enhancement method that significantly improves texture quality while preserving the appearance and geometry while fixing its degradations. Furthermore, Elevate3D operates in a view-by-view manner, alternating between texture and geometry refinement. Unlike previous methods that have largely overlooked geometry refinement, our framework leverages geometric cues from images refined with HFS-SDEdit by employing state-of-the-art monocular geometry predictors. This approach ensures detailed and accurate geometry that aligns seamlessly with the enhanced texture. Elevate3D outperforms recent competitors by achieving state-of-the-art quality in 3D model refinement, effectively addressing the scarcity of high-quality open-source 3D assets.
comment: Accepted to SIGGRAPH 2025. For the project page, see https://cg.postech.ac.kr/research/Elevate3D/
☆ Deep Equilibrium models for Poisson Imaging Inverse problems via Mirror Descent
Deep Equilibrium Models (DEQs) are implicit neural networks with fixed points, which have recently gained attention for learning image regularization functionals, particularly in settings involving Gaussian fidelities, where assumptions on the forward operator ensure contractiveness of standard (proximal) Gradient Descent operators. In this work, we extend the application of DEQs to Poisson inverse problems, where the data fidelity term is more appropriately modeled by the Kullback-Leibler divergence. To this end, we introduce a novel DEQ formulation based on Mirror Descent defined in terms of a tailored non-Euclidean geometry that naturally adapts with the structure of the data term. This enables the learning of neural regularizers within a principled training framework. We derive sufficient conditions to guarantee the convergence of the learned reconstruction scheme and propose computational strategies that enable both efficient training and fully parameter-free inference. Numerical experiments show that our method outperforms traditional model-based approaches and it is comparable to the performance of Bregman Plug-and-Play methods, while mitigating their typical drawbacks - namely, sensitivity to initialization and careful tuning of hyperparameters. The code is publicly available at https://github.com/christiandaniele/DEQ-MD.
☆ COLI: A Hierarchical Efficient Compressor for Large Images
The escalating adoption of high-resolution, large-field-of-view imagery amplifies the need for efficient compression methodologies. Conventional techniques frequently fail to preserve critical image details, while data-driven approaches exhibit limited generalizability. Implicit Neural Representations (INRs) present a promising alternative by learning continuous mappings from spatial coordinates to pixel intensities for individual images, thereby storing network weights rather than raw pixels and avoiding the generalization problem. However, INR-based compression of large images faces challenges including slow compression speed and suboptimal compression ratios. To address these limitations, we introduce COLI (Compressor for Large Images), a novel framework leveraging Neural Representations for Videos (NeRV). First, recognizing that INR-based compression constitutes a training process, we accelerate its convergence through a pretraining-finetuning paradigm, mixed-precision training, and reformulation of the sequential loss into a parallelizable objective. Second, capitalizing on INRs' transformation of image storage constraints into weight storage, we implement Hyper-Compression, a novel post-training technique to substantially enhance compression ratios while maintaining minimal output distortion. Evaluations across two medical imaging datasets demonstrate that COLI consistently achieves competitive or superior PSNR and SSIM metrics at significantly reduced bits per pixel (bpp), while accelerating NeRV training by up to 4 times.
☆ Implementing Adaptations for Vision AutoRegressive Model ICML 2025
Vision AutoRegressive model (VAR) was recently introduced as an alternative to Diffusion Models (DMs) in image generation domain. In this work we focus on its adaptations, which aim to fine-tune pre-trained models to perform specific downstream tasks, like medical data generation. While for DMs there exist many techniques, adaptations for VAR remain underexplored. Similarly, differentially private (DP) adaptations-ones that aim to preserve privacy of the adaptation data-have been extensively studied for DMs, while VAR lacks such solutions. In our work, we implement and benchmark many strategies for VAR, and compare them to state-of-the-art DM adaptation strategies. We observe that VAR outperforms DMs for non-DP adaptations, however, the performance of DP suffers, which necessitates further research in private adaptations for VAR. Code is available at https://github.com/sprintml/finetuning_var_dp.
comment: Accepted at DIG-BUGS: Data in Generative Models Workshop @ ICML 2025
☆ U-RWKV: Lightweight medical image segmentation with direction-adaptive RWKV MICCAI2025
Achieving equity in healthcare accessibility requires lightweight yet high-performance solutions for medical image segmentation, particularly in resource-limited settings. Existing methods like U-Net and its variants often suffer from limited global Effective Receptive Fields (ERFs), hindering their ability to capture long-range dependencies. To address this, we propose U-RWKV, a novel framework leveraging the Recurrent Weighted Key-Value(RWKV) architecture, which achieves efficient long-range modeling at O(N) computational cost. The framework introduces two key innovations: the Direction-Adaptive RWKV Module(DARM) and the Stage-Adaptive Squeeze-and-Excitation Module(SASE). DARM employs Dual-RWKV and QuadScan mechanisms to aggregate contextual cues across images, mitigating directional bias while preserving global context and maintaining high computational efficiency. SASE dynamically adapts its architecture to different feature extraction stages, balancing high-resolution detail preservation and semantic relationship capture. Experiments demonstrate that U-RWKV achieves state-of-the-art segmentation performance with high computational efficiency, offering a practical solution for democratizing advanced medical imaging technologies in resource-constrained environments. The code is available at https://github.com/hbyecoding/U-RWKV.
comment: Accepted by MICCAI2025
♻ ☆ Biomechanics-Guided Residual Approach to Generalizable Human Motion Generation and Estimation
Human pose, action, and motion generation are critical for applications in digital humans, character animation, and humanoid robotics. However, many existing methods struggle to produce physically plausible movements that are consistent with biomechanical principles. Although recent autoregressive and diffusion models deliver impressive visual quality, they often neglect key biodynamic features and fail to ensure physically realistic motions. Reinforcement Learning (RL) approaches can address these shortcomings but are highly dependent on simulation environments, limiting their generalizability. To overcome these challenges, we propose BioVAE, a biomechanics-aware framework with three core innovations: (1) integration of muscle electromyography (EMG) signals and kinematic features with acceleration constraints to enable physically plausible motion without simulations; (2) seamless coupling with diffusion models for stable end-to-end training; and (3) biomechanical priors that promote strong generalization across diverse motion generation and estimation tasks. Extensive experiments demonstrate that BioVAE achieves state-of-the-art performance on multiple benchmarks, bridging the gap between data-driven motion synthesis and biomechanical authenticity while setting new standards for physically accurate motion generation and pose estimation.
♻ ☆ Augmenting End-to-End Steering Angle Prediction with CAN Bus Data
In recent years, end to end steering prediction for autonomous vehicles has become a major area of research. The primary method for achieving end to end steering was to use computer vision models on a live feed of video data. However, to further increase accuracy, many companies have added data from light detection and ranging (LiDAR) and or radar sensors through sensor fusion. However, the addition of lasers and sensors comes at a high financial cost. In this paper, I address both of these issues by increasing the accuracy of the computer vision models without the increased cost of using LiDAR and or sensors. I achieved this by improving the accuracy of computer vision models by sensor fusing CAN bus data, a vehicle protocol, with video data. CAN bus data is a rich source of information about the vehicle's state, including its speed, steering angle, and acceleration. By fusing this data with video data, the accuracy of the computer vision model's predictions can be improved. When I trained the model without CAN bus data, I obtained an RMSE of 0.02492, while the model trained with the CAN bus data achieved an RMSE of 0.01970. This finding indicates that fusing CAN Bus data with video data can reduce the computer vision model's prediction error by 20% with some models decreasing the error by 80%.
comment: 5 pages
♻ ☆ QuEST: Low-bit Diffusion Model Quantization via Efficient Selective Finetuning ICCV 2025
The practical deployment of diffusion models is still hindered by the high memory and computational overhead. Although quantization paves a way for model compression and acceleration, existing methods face challenges in achieving low-bit quantization efficiently. In this paper, we identify imbalanced activation distributions as a primary source of quantization difficulty, and propose to adjust these distributions through weight finetuning to be more quantization-friendly. We provide both theoretical and empirical evidence supporting finetuning as a practical and reliable solution. Building on this approach, we further distinguish two critical types of quantized layers: those responsible for retaining essential temporal information and those particularly sensitive to bit-width reduction. By selectively finetuning these layers under both local and global supervision, we mitigate performance degradation while enhancing quantization efficiency. Our method demonstrates its efficacy across three high-resolution image generation tasks, obtaining state-of-the-art performance across multiple bit-width settings.
comment: ICCV 2025. Code is available at https://github.com/hatchetProject/QuEST
Information Retrieval 15
☆ Seq vs Seq: An Open Suite of Paired Encoders and Decoders
The large language model (LLM) community focuses almost exclusively on decoder-only language models, since they are easier to use for text generation. However, a large subset of the community still uses encoder-only models for tasks such as classification or retrieval. Previous work has attempted to compare these architectures, but is forced to make comparisons with models that have different numbers of parameters, training techniques, and datasets. We introduce the SOTA open-data Ettin suite of models: paired encoder-only and decoder-only models ranging from 17 million parameters to 1 billion, trained on up to 2 trillion tokens. Using the same recipe for both encoder-only and decoder-only models produces SOTA recipes in both categories for their respective sizes, beating ModernBERT as an encoder and Llama 3.2 and SmolLM2 as decoders. Like previous work, we find that encoder-only models excel at classification and retrieval tasks while decoders excel at generative tasks. However, we show that adapting a decoder model to encoder tasks (and vice versa) through continued training is subpar compared to using only the reverse objective (i.e. a 400M encoder outperforms a 1B decoder on MNLI, and vice versa for generative tasks). We open-source all artifacts of this study including training data, training order segmented by checkpoint, and 200+ checkpoints to allow future work to analyze or extend all aspects of training.
☆ From Chaos to Automation: Enabling the Use of Unstructured Data for Robotic Process Automation
The growing volume of unstructured data within organizations poses significant challenges for data analysis and process automation. Unstructured data, which lacks a predefined format, encompasses various forms such as emails, reports, and scans. It is estimated to constitute approximately 80% of enterprise data. Despite the valuable insights it can offer, extracting meaningful information from unstructured data is more complex compared to structured data. Robotic Process Automation (RPA) has gained popularity for automating repetitive tasks, improving efficiency, and reducing errors. However, RPA is traditionally reliant on structured data, limiting its application to processes involving unstructured documents. This study addresses this limitation by developing the UNstructured Document REtrieval SyStem (UNDRESS), a system that uses fuzzy regular expressions, techniques for natural language processing, and large language models to enable RPA platforms to effectively retrieve information from unstructured documents. The research involved the design and development of a prototype system, and its subsequent evaluation based on text extraction and information retrieval performance. The results demonstrate the effectiveness of UNDRESS in enhancing RPA capabilities for unstructured data, providing a significant advancement in the field. The findings suggest that this system could facilitate broader RPA adoption across processes traditionally hindered by unstructured data, thereby improving overall business process efficiency.
comment: Accepted at AUTOMATE 2025
☆ Aligned Query Expansion: Efficient Query Expansion for Information Retrieval through LLM Alignment
With the breakthroughs in large language models (LLMs), query generation techniques that expand documents and queries with related terms are becoming increasingly popular in the information retrieval field. Such techniques have been shown to improve the effectiveness of traditional lexical retrieval methods by dealing with the vocabulary mismatch problem. Recent work has found that generating queries with a greedy decoding strategy can produce sub-optimal queries, including hallucinations, and proposed to filter out queries before expansion. This `generate-then-filter' approach is costly, as it requires generating multiple queries and applying a relevance model to all of them and does not teach the LLM which of the generated queries is more effective for expansion. To overcome such limitations, we propose Aligned Query Expansion (AQE), a novel approach to enhance query expansion for passage retrieval in open-domain question answering. AQE leverages recent techniques in LLM alignment to fine-tune models for generating query expansions that directly optimize the effectiveness of the retrieval task, eliminating the need for additional filtering steps. This alignment ensures that queries are more relevant, reducing computational costs while improving retrieval effectiveness. Empirical evaluations show that AQE outperforms baseline models for query expansion in both in-domain and out-of-domain settings, demonstrating significant improvements in retrieval effectiveness.
☆ Unraveling the Biomarker Prospects of High-Altitude Diseases: Insights from Biomolecular Event Network Constructed using Text Mining
High-altitude diseases (HAD), encompassing acute mountain sickness (AMS), high-altitude cerebral edema (HACE), and high-altitude pulmonary edema (HAPE), are triggered by hypobaric hypoxia at elevations above 2,500 meters. These conditions pose significant health risks, yet the molecular mechanisms remain insufficiently understood. In this study, we developed a biomolecular event extraction pipeline integrating supervised machine learning with feature-based and multiscale Laplacian graph kernels to analyze 7,847 curated HAD-related abstracts from PubMed. We extracted over 150 unique biomolecular events including gene expression, regulation, binding, and localization and constructed a weighted, undirected biomolecular event network comprising 97 nodes and 153 edges. Using the PageRank algorithm, we prioritized key biomolecules based on their centrality within the event network. The top-ranked proteins included Erythropoietin (EPO) (0.0163), Vascular endothelial growth factor (VEGF) (0.0148), Hypoxia-inducible factor 1 (HIF-1) alpha (0.0136), Endothelial PAS Domain Protein 1 (EPAS1) and Angiotensin-Converting Enzyme (ACE) (0.0119), Egl nine homolog 1 (EGLN1), Endothelin 1 (ET-1), and 70 kilodalton heat shock protein (Hsp70)(0.0118), all of which play crucial roles in oxygen sensing, vascular remodeling, erythropoiesis, and blood pressure regulation. Subnetwork analysis revealed three major functional clusters centered on hypoxia response, inflammation, and stress adaptation pathways. Our integrative approach demonstrates the utility of large-scale text mining and graph-based analysis to uncover mechanistic insights and prioritize potential biomarkers for high-altitude disease.
☆ LLM-Driven Dual-Level Multi-Interest Modeling for Recommendation
Recently, much effort has been devoted to modeling users' multi-interests based on their behaviors or auxiliary signals. However, existing methods often rely on heuristic assumptions, e.g., co-occurring items indicate the same interest of users, failing to capture user multi-interests aligning with real-world scenarios. While large language models (LLMs) show significant potential for multi-interest analysis due to their extensive knowledge and powerful reasoning capabilities, two key challenges remain. First, the granularity of LLM-driven multi-interests is agnostic, possibly leading to overly fine or coarse interest grouping. Second, individual user analysis provides limited insights due to the data sparsity issue. In this paper, we propose an LLM-driven dual-level multi-interest modeling framework for more effective recommendation. At the user-individual level, we exploit LLMs to flexibly allocate items engaged by users into different semantic clusters, indicating their diverse and distinct interests. To alleviate the agnostic generation of LLMs, we adaptively assign these semantic clusters to users' collaborative multi-interests learned from global user-item interactions, allowing the granularity to be automatically adjusted according to the user's behaviors using an alignment module. To alleviate the limited insights derived from individual users' behaviors, at the user-crowd level, we propose aggregating user cliques into synthesized users with rich behaviors for more comprehensive LLM-driven multi-interest analysis. We formulate a max covering problem to ensure the compactness and representativeness of synthesized users' behaviors, and then conduct contrastive learning based on their LLM-driven multi-interests to disentangle item representations among different interests. Experiments on real-world datasets show the superiority of our approach against state-of-the-art methods.
comment: 10 pages, 5 figures
☆ AI Wizards at CheckThat! 2025: Enhancing Transformer-Based Embeddings with Sentiment for Subjectivity Detection in News Articles
This paper presents AI Wizards' participation in the CLEF 2025 CheckThat! Lab Task 1: Subjectivity Detection in News Articles, classifying sentences as subjective/objective in monolingual, multilingual, and zero-shot settings. Training/development datasets were provided for Arabic, German, English, Italian, and Bulgarian; final evaluation included additional unseen languages (e.g., Greek, Romanian, Polish, Ukrainian) to assess generalization. Our primary strategy enhanced transformer-based classifiers by integrating sentiment scores, derived from an auxiliary model, with sentence representations, aiming to improve upon standard fine-tuning. We explored this sentiment-augmented architecture with mDeBERTaV3-base, ModernBERT-base (English), and Llama3.2-1B. To address class imbalance, prevalent across languages, we employed decision threshold calibration optimized on the development set. Our experiments show sentiment feature integration significantly boosts performance, especially subjective F1 score. This framework led to high rankings, notably 1st for Greek (Macro F1 = 0.51).
comment: 14 pages, 6 figures, accepted at CLEF 2025 CheckThat! Lab
☆ An Empirical Study of Multi-Agent RAG for Real-World University Admissions Counseling
This paper presents MARAUS (Multi-Agent and Retrieval-Augmented University Admission System), a real-world deployment of a conversational AI platform for higher education admissions counseling in Vietnam. While large language models (LLMs) offer potential for automating advisory tasks, most existing solutions remain limited to prototypes or synthetic benchmarks. MARAUS addresses this gap by combining hybrid retrieval, multi-agent orchestration, and LLM-based generation into a system tailored for real-world university admissions. In collaboration with the University of Transport Technology (UTT) in Hanoi, we conducted a two-phase study involving technical development and real-world evaluation. MARAUS processed over 6,000 actual user interactions, spanning six categories of queries. Results show substantial improvements over LLM-only baselines: on average 92 percent accuracy, hallucination rates reduced from 15 precent to 1.45 percent, and average response times below 4 seconds. The system operated cost-effectively, with a two-week deployment cost of 11.58 USD using GPT-4o mini. This work provides actionable insights for the deployment of agentic RAG systems in low-resource educational settings.
♻ ☆ Alleviating User-Sensitive bias with Fair Generative Sequential Recommendation Model
Recommendation fairness has recently attracted much attention. In the real world, recommendation systems are driven by user behavior, and since users with the same sensitive feature (e.g., gender and age) tend to have the same patterns, recommendation models can easily capture the strong correlation preference of sensitive features and thus cause recommendation unfairness. Diffusion model (DM) as a new generative model paradigm has achieved great success in recommendation systems. DM's ability to model uncertainty and represent diversity, and its modeling mechanism has a high degree of adaptability with the real-world recommendation process with bias. Therefore, we use DM to effectively model the fairness of recommendation and enhance the diversity. This paper proposes a FairGENerative sequential Recommendation model based on DM, FairGENRec. In the training phase, we inject random noise into the original distribution under the guidance of the sensitive feature recognition model, and a sequential denoise model is designed for the reverse reconstruction of items. Simultaneously, recommendation fairness modeling is completed by injecting multi-interests representational information that eliminates the bias of sensitive user features into the generated results. In the inference phase, the model obtains the noise in the form of noise addition by using the history interactions which is followed by reverse iteration to reconstruct the target item representation. Finally, our extensive experiments on three datasets demonstrate the dual enhancement effect of FairGENRec on accuracy and fairness, while the statistical analysis of the cases visualizes the degree of improvement on the fairness of the recommendation.
♻ ☆ Are AI Agents interacting with Online Ads?
As AI-driven agents become increasingly integrated into the digital ecosystem, they reshape how online advertising is perceived and processed. Particularly in the travel and hotel booking sector, these autonomous systems influence the effectiveness of traditional advertising formats. While visual cues and emotional appeals sway human users, AI agents prioritize structured data such as price, availability, and specifications. This study examines how different AI agents interact with online advertising, whether they incorporate ads into their decision-making processes, and which ad formats prove most effective. We analyze interaction patterns, click behavior, and decision-making strategies through experiments with multimodal language models such as OpenAI GPT-4o, Anthropic Claude 3.7 Sonnet, and Google Gemini 2.0 Flash. Our findings reveal that AI agents neither ignore nor systematically avoid advertisements but instead favor certain features-particularly keywords and structured data. These insights have significant implications for the future design of advertising strategies in AI-dominated digital environments.
♻ ☆ RAG-R1 : Incentivize the Search and Reasoning Capabilities of LLMs through Multi-query Parallelism
Large Language Models (LLMs) have demonstrated remarkable capabilities across various tasks, while they remain prone to generating hallucinated or outdated responses due to their static internal knowledge. Recent advancements in Retrieval-Augmented Generation (RAG) methods have explored enhancing models' search and reasoning capabilities through reinforcement learning (RL). Although these methods demonstrate promising results, they face challenges in training stability and encounter issues such as substantial inference time and restricted capabilities due to the single-query mode. In this paper, we propose RAG-R1, a novel training framework designed to enable LLMs to adaptively leverage internal and external knowledge during the reasoning process. We further expand the generation and retrieval processes within the framework from single-query mode to multi-query parallelism, aimed at reducing inference time and enhancing the model's capabilities. Extensive experiments on seven question-answering benchmarks demonstrate that our method outperforms the strongest baseline by up to 13.2% and decreases inference time by 11.1%.
♻ ☆ Evaluating Multimodal Large Language Models on Educational Textbook Question Answering
Multimodal large language models (MLLMs) have shown success in vision-language tasks, but their ability to reason over complex educational materials remains largely untested. This work presents the first evaluation of state-of-the-art MLLMs, including LLaVA-1.5 and LLaMA 3.2-Vision, on the textbook question answering (TQA) task using the CK12-QA dataset. We introduce a multimodal retrieval-augmented generation (RAG) pipeline to simulate real-world learning by providing relevant lesson paragraphs and diagrams as context. Our zero-shot experiments reveal a critical trade-off: while retrieved context improves LLaVA's performance on text-based questions, it significantly degrades the accuracy of the more powerful LLaMA 3.2-Vision on diagram-based tasks, dropping its validation accuracy from 74.07% to 25.93%. We term this statistically significant phenomenon "catastrophic context interference." Furthermore, fine-tuning highlights architectural differences: LLaMA 3.2-Vision's performance improves to 71.16% on the test set, demonstrating its capacity to learn multimodal integration, whereas LLaVA's performance declines, indicating challenges with generalization. Our results underscore the challenges MLLMs face in modality prioritization and context integration, providing a benchmark and pointing to key directions for developing more robust AI-driven educational tools.
comment: 8 Pages
♻ ☆ Big data searching using words
Big data analytics is one of the most promising areas of new research and development in computer science, enterprises, e-commerce, and defense. For many organizations, big data is considered one of their most important strategic assets. This explosive growth has made it necessary to develop effective techniques for examining and analyzing big data from mathematical perspectives. Among various methods of analyzing big data, topological data analysis (TDA) is now considered one of the useful tools. However, there is no fundamental concept related to the topological structure in big data. In this paper, we present fundamental concepts related to the neighborhood structures of words in big data search, laying the groundwork for developing topological frameworks for big data in the future. We also introduce the notion of big data primal within the context of big data search and explore how neighborhood structures, combined with the Jaccard similarity coefficient, can be utilized to detect anomalies in search behavior.
comment: accepted for publication, Acceptance month: July, 2025
♻ ☆ CART: A Generative Cross-Modal Retrieval Framework with Coarse-To-Fine Semantic Modeling ACL 2025
Cross-modal retrieval aims to search for instances, which are semantically related to the query through the interaction of different modal data. Traditional solutions utilize a single-tower or dual-tower framework to explicitly compute the score between queries and candidates, which is challenged by training cost and inference latency with large-scale data. Inspired by the remarkable performance and efficiency of generative models, we propose a generative cross-modal retrieval framework (CART) based on coarse-to-fine semantic modeling, which assigns identifiers to each candidate and treats the generating identifier as the retrieval target. Specifically, we explore an effective coarse-to-fine scheme, combining K-Means and RQ-VAE to discretize multimodal data into token sequences that support autoregressive generation. Further, considering the lack of explicit interaction between queries and candidates, we propose a feature fusion strategy to align their semantics. Extensive experiments demonstrate the effectiveness of the strategies in the CART, achieving excellent results in both retrieval performance and efficiency.
comment: ACL 2025 Main
♻ ☆ Repeated Padding+: Simple yet Effective Data Augmentation Plugin for Sequential Recommendation RecSys 2024
Sequential recommendation aims to provide users with personalized suggestions based on their historical interactions. When training sequential models, padding is a widely adopted technique for two main reasons: 1) The vast majority of models can only handle fixed-length sequences; 2) Batching-based training needs to ensure that the sequences in each batch have the same length. The special value \emph{0} is usually used as the padding content, which does not contain the actual information and is ignored in the model calculations. This common-sense padding strategy leads us to a problem that has never been explored before: Can we fully utilize this idle input space by padding other content to further improve model performance and training efficiency? In this work, we propose a simple yet effective padding method called Repeated Padding+ (RepPad+). Specifically, we use the original interaction sequences as the padding content and fill it to the padding positions during model training. This operation can be performed a finite number of times or repeated until the input sequences' length reaches the maximum limit. For those sequences that can not pad full original data, we draw inspiration from the Sliding Windows strategy and intercept consecutive subsequences to fill in the idle space. Our RepPad+ can be viewed as a sequence-level data augmentation strategy. Unlike most existing works, our method contains no trainable parameters or hyperparameters and is a plug-and-play data augmentation operation. Extensive experiments on various categories of sequential models and seven real-world datasets demonstrate the effectiveness and efficiency of our approach. The average recommendation performance improvement is up to 84.11% on GRU4Rec and 35.34% on SASRec. We also provide in-depth analysis and explanation of what makes RepPad+ effective from multiple perspectives.
comment: v1: Initial Version, v2:Accepted by RecSys 2024, v3:Extended Version Under Review
♻ ☆ A quantum semantic framework for natural language processing
Semantic degeneracy represents a fundamental property of natural language that extends beyond simple polysemy to encompass the combinatorial explosion of potential interpretations that emerges as semantic expressions increase in complexity. In this work, we argue this property imposes fundamental limitations on Large Language Models (LLMs) and other modern NLP systems, precisely because they operate within natural language itself. Using Kolmogorov complexity, we demonstrate that as an expression's complexity grows, the amount of contextual information required to reliably resolve its ambiguity explodes combinatorially. The computational intractability of recovering a single intended meaning for complex or ambiguous text therefore suggests that the classical view that linguistic forms possess intrinsic meaning in and of themselves is conceptually inadequate. We argue instead that meaning is dynamically actualized through an observer-dependent interpretive act, a process whose non-deterministic nature is most appropriately described by a non-classical, quantum-like logic. To test this hypothesis, we conducted a semantic Bell inequality test using diverse LLM agents. Our experiments yielded average CHSH expectation values from 1.2 to 2.8, with several runs producing values (e.g., 2.3-2.4) in significant violation of the classical boundary ($|S|\leq2$), demonstrating that linguistic interpretation under ambiguity can exhibit non-classical contextuality, consistent with results from human cognition experiments. These results inherently imply that classical frequentist-based analytical approaches for natural language are necessarily lossy. Instead, we propose that Bayesian-style repeated sampling approaches can provide more practically useful and appropriate characterizations of linguistic meaning in context.
comment: 12 pages, 2 figures, accepted submission to Quantum AI and NLP 2025
Machine Learning 150
☆ Streaming 4D Visual Geometry Transformer
Perceiving and reconstructing 4D spatial-temporal geometry from videos is a fundamental yet challenging computer vision task. To facilitate interactive and real-time applications, we propose a streaming 4D visual geometry transformer that shares a similar philosophy with autoregressive large language models. We explore a simple and efficient design and employ a causal transformer architecture to process the input sequence in an online manner. We use temporal causal attention and cache the historical keys and values as implicit memory to enable efficient streaming long-term 4D reconstruction. This design can handle real-time 4D reconstruction by incrementally integrating historical information while maintaining high-quality spatial consistency. For efficient training, we propose to distill knowledge from the dense bidirectional visual geometry grounded transformer (VGGT) to our causal model. For inference, our model supports the migration of optimized efficient attention operator (e.g., FlashAttention) from the field of large language models. Extensive experiments on various 4D geometry perception benchmarks demonstrate that our model increases the inference speed in online scenarios while maintaining competitive performance, paving the way for scalable and interactive 4D vision systems. Code is available at: https://github.com/wzzheng/StreamVGGT.
comment: Code is available at: https://github.com/wzzheng/StreamVGGT
☆ Canonical Bayesian Linear System Identification
Standard Bayesian approaches for linear time-invariant (LTI) system identification are hindered by parameter non-identifiability; the resulting complex, multi-modal posteriors make inference inefficient and impractical. We solve this problem by embedding canonical forms of LTI systems within the Bayesian framework. We rigorously establish that inference in these minimal parameterizations fully captures all invariant system dynamics (e.g., transfer functions, eigenvalues, predictive distributions of system outputs) while resolving identifiability. This approach unlocks the use of meaningful, structure-aware priors (e.g., enforcing stability via eigenvalues) and ensures conditions for a Bernstein--von Mises theorem -- a link between Bayesian and frequentist large-sample asymptotics that is broken in standard forms. Extensive simulations with modern MCMC methods highlight advantages over standard parameterizations: canonical forms achieve higher computational efficiency, generate interpretable and well-behaved posteriors, and provide robust uncertainty estimates, particularly from limited data.
comment: 46 pages, 9 figures
☆ Langevin Flows for Modeling Neural Latent Dynamics
Neural populations exhibit latent dynamical structures that drive time-evolving spiking activities, motivating the search for models that capture both intrinsic network dynamics and external unobserved influences. In this work, we introduce LangevinFlow, a sequential Variational Auto-Encoder where the time evolution of latent variables is governed by the underdamped Langevin equation. Our approach incorporates physical priors -- such as inertia, damping, a learned potential function, and stochastic forces -- to represent both autonomous and non-autonomous processes in neural systems. Crucially, the potential function is parameterized as a network of locally coupled oscillators, biasing the model toward oscillatory and flow-like behaviors observed in biological neural populations. Our model features a recurrent encoder, a one-layer Transformer decoder, and Langevin dynamics in the latent space. Empirically, our method outperforms state-of-the-art baselines on synthetic neural populations generated by a Lorenz attractor, closely matching ground-truth firing rates. On the Neural Latents Benchmark (NLB), the model achieves superior held-out neuron likelihoods (bits per spike) and forward prediction accuracy across four challenging datasets. It also matches or surpasses alternative methods in decoding behavioral metrics such as hand velocity. Overall, this work introduces a flexible, physics-inspired, high-performing framework for modeling complex neural population dynamics and their unobserved influences.
comment: Full version of the Cognitive Computational Neuroscience (CCN) 2025 poster
☆ CATVis: Context-Aware Thought Visualization MICCAI 2025
EEG-based brain-computer interfaces (BCIs) have shown promise in various applications, such as motor imagery and cognitive state monitoring. However, decoding visual representations from EEG signals remains a significant challenge due to their complex and noisy nature. We thus propose a novel 5-stage framework for decoding visual representations from EEG signals: (1) an EEG encoder for concept classification, (2) cross-modal alignment of EEG and text embeddings in CLIP feature space, (3) caption refinement via re-ranking, (4) weighted interpolation of concept and caption embeddings for richer semantics, and (5) image generation using a pre-trained Stable Diffusion model. We enable context-aware EEG-to-image generation through cross-modal alignment and re-ranking. Experimental results demonstrate that our method generates high-quality images aligned with visual stimuli, outperforming SOTA approaches by 13.43% in Classification Accuracy, 15.21% in Generation Accuracy and reducing Fr\'echet Inception Distance by 36.61%, indicating superior semantic alignment and image quality.
comment: Accepted at MICCAI 2025. This is the submitted version prior to peer review. The final Version of Record will appear in the MICCAI 2025 proceedings (Springer LNCS)
☆ AirLLM: Diffusion Policy-based Adaptive LoRA for Remote Fine-Tuning of LLM over the Air
Operating Large Language Models (LLMs) on edge devices is increasingly challenged by limited communication bandwidth and strained computational and memory costs. Thus, cloud-assisted remote fine-tuning becomes indispensable. Nevertheless, existing Low-Rank Adaptation (LoRA) approaches typically employ fixed or heuristic rank configurations, and the subsequent over-the-air transmission of all LoRA parameters could be rather inefficient. To address this limitation, we develop AirLLM, a hierarchical diffusion policy framework for communication-aware LoRA adaptation. Specifically, AirLLM models the rank configuration as a structured action vector that spans all LoRA-inserted projections. To solve the underlying high-dimensional sequential decision-making problem, a Proximal Policy Optimization (PPO) agent generates coarse-grained decisions by jointly observing wireless states and linguistic complexity, which are then refined via Denoising Diffusion Implicit Models (DDIM) to produce high-resolution, task- and channel-adaptive rank vectors. The two modules are optimized alternatively, with the DDIM trained under the Classifier-Free Guidance (CFG) paradigm to maintain alignment with PPO rewards. Experiments under varying signal-to-noise ratios demonstrate that AirLLM consistently enhances fine-tuning performance while significantly reducing transmission costs, highlighting the effectiveness of reinforcement-driven, diffusion-refined rank adaptation for scalable and efficient remote fine-tuning over the air.
comment: 11 pages, 8 figures
☆ Elk: Exploring the Efficiency of Inter-core Connected AI Chips with Deep Learning Compiler Techniques MICRO'25
To meet the increasing demand of deep learning (DL) models, AI chips are employing both off-chip memory (e.g., HBM) and high-bandwidth low-latency interconnect for direct inter-core data exchange. However, it is not easy to explore the efficiency of these inter-core connected AI (ICCA) chips, due to a fundamental tussle among compute (per-core execution), communication (inter-core data exchange), and I/O (off-chip data access). In this paper, we develop Elk, a DL compiler framework to maximize the efficiency of ICCA chips by jointly trading off all the three performance factors discussed above. Elk structures these performance factors into configurable parameters and forms a global trade-off space in the DL compiler. To systematically explore this space and maximize overall efficiency, Elk employs a new inductive operator scheduling policy and a cost-aware on-chip memory allocation algorithm. It generates globally optimized execution plans that best overlap off-chip data loading and on-chip execution. To examine the efficiency of Elk, we build a full-fledged emulator based on a real ICCA chip IPU-POD4, and an ICCA chip simulator for sensitivity analysis with different interconnect network topologies. Elk achieves 94% of the ideal roofline performance of ICCA chips on average, showing the benefits of supporting large DL models on ICCA chips. We also show Elk's capability of enabling architecture design space exploration for new ICCA chip development.
comment: This paper is accepted at the 58th IEEE/ACM International Symposium on Microarchitecture (MICRO'25)
☆ Exploring the robustness of TractOracle methods in RL-based tractography
Tractography algorithms leverage diffusion MRI to reconstruct the fibrous architecture of the brain's white matter. Among machine learning approaches, reinforcement learning (RL) has emerged as a promising framework for tractography, outperforming traditional methods in several key aspects. TractOracle-RL, a recent RL-based approach, reduces false positives by incorporating anatomical priors into the training process via a reward-based mechanism. In this paper, we investigate four extensions of the original TractOracle-RL framework by integrating recent advances in RL, and we evaluate their performance across five diverse diffusion MRI datasets. Results demonstrate that combining an oracle with the RL framework consistently leads to robust and reliable tractography, regardless of the specific method or dataset used. We also introduce a novel RL training scheme called Iterative Reward Training (IRT), inspired by the Reinforcement Learning from Human Feedback (RLHF) paradigm. Instead of relying on human input, IRT leverages bundle filtering methods to iteratively refine the oracle's guidance throughout training. Experimental results show that RL methods trained with oracle feedback significantly outperform widely used tractography techniques in terms of accuracy and anatomical validity.
comment: 38 pages, 8 figures. Submitted to Medical Image Analysis
Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety
AI systems that "think" in human language offer a unique opportunity for AI safety: we can monitor their chains of thought (CoT) for the intent to misbehave. Like all other known AI oversight methods, CoT monitoring is imperfect and allows some misbehavior to go unnoticed. Nevertheless, it shows promise and we recommend further research into CoT monitorability and investment in CoT monitoring alongside existing safety methods. Because CoT monitorability may be fragile, we recommend that frontier model developers consider the impact of development decisions on CoT monitorability.
☆ D3FL: Data Distribution and Detrending for Robust Federated Learning in Non-linear Time-series Data
With advancements in computing and communication technologies, the Internet of Things (IoT) has seen significant growth. IoT devices typically collect data from various sensors, such as temperature, humidity, and energy meters. Much of this data is temporal in nature. Traditionally, data from IoT devices is centralized for analysis, but this approach introduces delays and increased communication costs. Federated learning (FL) has emerged as an effective alternative, allowing for model training across distributed devices without the need to centralize data. In many applications, such as smart home energy and environmental monitoring, the data collected by IoT devices across different locations can exhibit significant variation in trends and seasonal patterns. Accurately forecasting such non-stationary, non-linear time-series data is crucial for applications like energy consumption estimation and weather forecasting. However, these data variations can severely impact prediction accuracy. The key contributions of this paper are: (1) Investigating how non-linear, non-stationary time-series data distributions, like generalized extreme value (gen-extreme) and log norm distributions, affect FL performance. (2) Analyzing how different detrending techniques for non-linear time-series data influence the forecasting model's performance in a FL setup. We generated several synthetic time-series datasets using non-linear data distributions and trained an LSTM-based forecasting model using both centralized and FL approaches. Additionally, we evaluated the impact of detrending on real-world datasets with non-linear time-series data distributions. Our experimental results show that: (1) FL performs worse than centralized approaches when dealing with non-linear data distributions. (2) The use of appropriate detrending techniques improves FL performance, reducing loss across different data distributions.
comment: Preprint of paper to appear in the proceedings of IEEE INTERNATIONAL CONFERENCE ON EDGE COMPUTING & COMMUNICATIONS EDGE 2025
☆ LRMR: LLM-Driven Relational Multi-node Ranking for Lymph Node Metastasis Assessment in Rectal Cancer
Accurate preoperative assessment of lymph node (LN) metastasis in rectal cancer guides treatment decisions, yet conventional MRI evaluation based on morphological criteria shows limited diagnostic performance. While some artificial intelligence models have been developed, they often operate as black boxes, lacking the interpretability needed for clinical trust. Moreover, these models typically evaluate nodes in isolation, overlooking the patient-level context. To address these limitations, we introduce LRMR, an LLM-Driven Relational Multi-node Ranking framework. This approach reframes the diagnostic task from a direct classification problem into a structured reasoning and ranking process. The LRMR framework operates in two stages. First, a multimodal large language model (LLM) analyzes a composite montage image of all LNs from a patient, generating a structured report that details ten distinct radiological features. Second, a text-based LLM performs pairwise comparisons of these reports between different patients, establishing a relative risk ranking based on the severity and number of adverse features. We evaluated our method on a retrospective cohort of 117 rectal cancer patients. LRMR achieved an area under the curve (AUC) of 0.7917 and an F1-score of 0.7200, outperforming a range of deep learning baselines, including ResNet50 (AUC 0.7708). Ablation studies confirmed the value of our two main contributions: removing the relational ranking stage or the structured prompting stage led to a significant performance drop, with AUCs falling to 0.6875 and 0.6458, respectively. Our work demonstrates that decoupling visual perception from cognitive reasoning through a two-stage LLM framework offers a powerful, interpretable, and effective new paradigm for assessing lymph node metastasis in rectal cancer.
☆ Implementing Adaptations for Vision AutoRegressive Model ICML 2025
Vision AutoRegressive model (VAR) was recently introduced as an alternative to Diffusion Models (DMs) in image generation domain. In this work we focus on its adaptations, which aim to fine-tune pre-trained models to perform specific downstream tasks, like medical data generation. While for DMs there exist many techniques, adaptations for VAR remain underexplored. Similarly, differentially private (DP) adaptations-ones that aim to preserve privacy of the adaptation data-have been extensively studied for DMs, while VAR lacks such solutions. In our work, we implement and benchmark many strategies for VAR, and compare them to state-of-the-art DM adaptation strategies. We observe that VAR outperforms DMs for non-DP adaptations, however, the performance of DP suffers, which necessitates further research in private adaptations for VAR. Code is available at https://github.com/sprintml/finetuning_var_dp.
comment: Accepted at DIG-BUGS: Data in Generative Models Workshop @ ICML 2025
☆ Data Augmentation in Time Series Forecasting through Inverted Framework
Currently, iTransformer is one of the most popular and effective models for multivariate time series (MTS) forecasting. Thanks to its inverted framework, iTransformer effectively captures multivariate correlation. However, the inverted framework still has some limitations. It diminishes temporal interdependency information, and introduces noise in cases of nonsignificant variable correlation. To address these limitations, we introduce a novel data augmentation method on inverted framework, called DAIF. Unlike previous data augmentation methods, DAIF stands out as the first real-time augmentation specifically designed for the inverted framework in MTS forecasting. We first define the structure of the inverted sequence-to-sequence framework, then propose two different DAIF strategies, Frequency Filtering and Cross-variation Patching to address the existing challenges of the inverted framework. Experiments across multiple datasets and inverted models have demonstrated the effectiveness of our DAIF.
☆ Toward Improving fNIRS Classification: A Study on Activation Functions in Deep Neural Architectures
Activation functions are critical to the performance of deep neural networks, particularly in domains such as functional near-infrared spectroscopy (fNIRS), where nonlinearity, low signal-to-noise ratio (SNR), and signal variability poses significant challenges to model accuracy. However, the impact of activation functions on deep learning (DL) performance in the fNIRS domain remains underexplored and lacks systematic investigation in the current literature. This study evaluates a range of conventional and field-specific activation functions for fNIRS classification tasks using multiple deep learning architectures, including the domain-specific fNIRSNet, AbsoluteNet, MDNN, and shallowConvNet (as the baseline), all tested on a single dataset recorded during an auditory task. To ensure fair a comparison, all networks were trained and tested using standardized preprocessing and consistent training parameters. The results show that symmetrical activation functions such as Tanh and the Absolute value function Abs(x) can outperform commonly used functions like the Rectified Linear Unit (ReLU), depending on the architecture. Additionally, a focused analysis of the role of symmetry was conducted using a Modified Absolute Function (MAF), with results further supporting the effectiveness of symmetrical activation functions on performance gains. These findings underscore the importance of selecting proper activation functions that align with the signal characteristics of fNIRS data.
☆ FLsim: A Modular and Library-Agnostic Simulation Framework for Federated Learning
Federated Learning (FL) has undergone significant development since its inception in 2016, advancing from basic algorithms to complex methodologies tailored to address diverse challenges and use cases. However, research and benchmarking of novel FL techniques against a plethora of established state-of-the-art solutions remain challenging. To streamline this process, we introduce FLsim, a comprehensive FL simulation framework designed to meet the diverse requirements of FL workflows in the literature. FLsim is characterized by its modularity, scalability, resource efficiency, and controlled reproducibility of experimental outcomes. Its easy to use interface allows users to specify customized FL requirements through job configuration, which supports: (a) customized data distributions, ranging from non-independent and identically distributed (non-iid) data to independent and identically distributed (iid) data, (b) selection of local learning algorithms according to user preferences, with complete agnosticism to ML libraries, (c) choice of network topology illustrating communication patterns among nodes, (d) definition of model aggregation and consensus algorithms, and (e) pluggable blockchain support for enhanced robustness. Through a series of experimental evaluations, we demonstrate the effectiveness and versatility of FLsim in simulating a diverse range of state-of-the-art FL experiments. We envisage that FLsim would mark a significant advancement in FL simulation frameworks, offering unprecedented flexibility and functionality for researchers and practitioners alike.
☆ Better Regret Rates in Bilateral Trade via Sublinear Budget Violation
Bilateral trade is a central problem in algorithmic economics, and recent work has explored how to design trading mechanisms using no-regret learning algorithms. However, no-regret learning is impossible when budget balance has to be enforced at each time step. Bernasconi et al. [Ber+24] show how this impossibility can be circumvented by relaxing the budget balance constraint to hold only globally over all time steps. In particular, they design an algorithm achieving regret of the order of $\tilde O(T^{3/4})$ and provide a lower bound of $\Omega(T^{5/7})$. In this work, we interpolate between these two extremes by studying how the optimal regret rate varies with the allowed violation of the global budget balance constraint. Specifically, we design an algorithm that, by violating the constraint by at most $T^{\beta}$ for any given $\beta \in [\frac{3}{4}, \frac{6}{7}]$, attains regret $\tilde O(T^{1 - \beta/3})$. We complement this result with a matching lower bound, thus fully characterizing the trade-off between regret and budget violation. Our results show that both the $\tilde O(T^{3/4})$ upper bound in the global budget balance case and the $\Omega(T^{5/7})$ lower bound under unconstrained budget balance violation obtained by Bernasconi et al. [Ber+24] are tight.
☆ Seq vs Seq: An Open Suite of Paired Encoders and Decoders
The large language model (LLM) community focuses almost exclusively on decoder-only language models, since they are easier to use for text generation. However, a large subset of the community still uses encoder-only models for tasks such as classification or retrieval. Previous work has attempted to compare these architectures, but is forced to make comparisons with models that have different numbers of parameters, training techniques, and datasets. We introduce the SOTA open-data Ettin suite of models: paired encoder-only and decoder-only models ranging from 17 million parameters to 1 billion, trained on up to 2 trillion tokens. Using the same recipe for both encoder-only and decoder-only models produces SOTA recipes in both categories for their respective sizes, beating ModernBERT as an encoder and Llama 3.2 and SmolLM2 as decoders. Like previous work, we find that encoder-only models excel at classification and retrieval tasks while decoders excel at generative tasks. However, we show that adapting a decoder model to encoder tasks (and vice versa) through continued training is subpar compared to using only the reverse objective (i.e. a 400M encoder outperforms a 1B decoder on MNLI, and vice versa for generative tasks). We open-source all artifacts of this study including training data, training order segmented by checkpoint, and 200+ checkpoints to allow future work to analyze or extend all aspects of training.
☆ Robust-Multi-Task Gradient Boosting
Multi-task learning (MTL) has shown effectiveness in exploiting shared information across tasks to improve generalization. MTL assumes tasks share similarities that can improve performance. In addition, boosting algorithms have demonstrated exceptional performance across diverse learning problems, primarily due to their ability to focus on hard-to-learn instances and iteratively reduce residual errors. This makes them a promising approach for learning multi-task problems. However, real-world MTL scenarios often involve tasks that are not well-aligned (known as outlier or adversarial tasks), which do not share beneficial similarities with others and can, in fact, deteriorate the performance of the overall model. To overcome this challenge, we propose Robust-Multi-Task Gradient Boosting (R-MTGB), a novel boosting framework that explicitly models and adapts to task heterogeneity during training. R-MTGB structures the learning process into three sequential blocks: (1) learning shared patterns, (2) partitioning tasks into outliers and non-outliers with regularized parameters, and (3) fine-tuning task-specific predictors. This architecture enables R-MTGB to automatically detect and penalize outlier tasks while promoting effective knowledge transfer among related tasks. Our method integrates these mechanisms seamlessly within gradient boosting, allowing robust handling of noisy or adversarial tasks without sacrificing accuracy. Extensive experiments on both synthetic benchmarks and real-world datasets demonstrate that our approach successfully isolates outliers, transfers knowledge, and consistently reduces prediction errors for each task individually, and achieves overall performance gains across all tasks. These results highlight robustness, adaptability, and reliable convergence of R-MTGB in challenging MTL environments.
☆ Stochastic Entanglement Configuration for Constructive Entanglement Topologies in Quantum Machine Learning with Application to Cardiac MRI
Efficient entanglement strategies are essential for advancing variational quantum circuits (VQCs) for quantum machine learning (QML). However, most current approaches use fixed entanglement topologies that are not adaptive to task requirements, limiting potential gains over classical models. We introduce a novel stochastic entanglement configuration method that systematically generates diverse entanglement topologies to identify a subspace of constructive entanglement configurations, defined as entanglement topologies that boost hybrid model performance (e.g., classification accuracy) beyond classical baselines. Each configuration is encoded as a stochastic binary matrix, denoting directed entanglement between qubits. This enables scalable exploration of the hyperspace of candidate entanglement topologies using entanglement density and per-qubit constraints as key metrics. We define unconstrained and constrained sampling modes, controlling entanglement per qubit. Using our method, 400 stochastic configurations were generated and evaluated in a hybrid QML for cardiac MRI disease classification. We identified 64 (16%) novel constructive entanglement configurations that consistently outperformed the classical baseline. Ensemble aggregation of top-performing configurations achieved ~0.92 classification accuracy, exceeding the classical model (~0.87) by over 5%. Compared to four conventional topologies (ring, nearest neighbor, no entanglement, fully entangled), none surpassed the classical baseline (maximum accuracy ~0.82), while our configurations delivered up to ~20% higher accuracy. Thus, highlighting the robustness and generalizability of the identified constructive entanglements.
comment: Accepted for publication at IEEE International Conference on Quantum Computing and Engineering (QCE) 2025
☆ A Neural Network Model of Complementary Learning Systems: Pattern Separation and Completion for Continual Learning
Learning new information without forgetting prior knowledge is central to human intelligence. In contrast, neural network models suffer from catastrophic forgetting: a significant degradation in performance on previously learned tasks when acquiring new information. The Complementary Learning Systems (CLS) theory offers an explanation for this human ability, proposing that the brain has distinct systems for pattern separation (encoding distinct memories) and pattern completion (retrieving complete memories from partial cues). To capture these complementary functions, we leverage the representational generalization capabilities of variational autoencoders (VAEs) and the robust memory storage properties of Modern Hopfield networks (MHNs), combining them into a neurally plausible continual learning model. We evaluate this model on the Split-MNIST task, a popular continual learning benchmark, and achieve close to state-of-the-art accuracy (~90%), substantially reducing forgetting. Representational analyses empirically confirm the functional dissociation: the VAE underwrites pattern completion, while the MHN drives pattern separation. By capturing pattern separation and completion in scalable architectures, our work provides a functional template for modeling memory consolidation, generalization, and continual learning in both biological and artificial systems.
comment: Accepted to CogSci 2025. 7 pages, 7 figures
☆ From Kinetic Theory to AI: a Rediscovery of High-Dimensional Divergences and Their Properties
Selecting an appropriate divergence measure is a critical aspect of machine learning, as it directly impacts model performance. Among the most widely used, we find the Kullback-Leibler (KL) divergence, originally introduced in kinetic theory as a measure of relative entropy between probability distributions. Just as in machine learning, the ability to quantify the proximity of probability distributions plays a central role in kinetic theory. In this paper, we present a comparative review of divergence measures rooted in kinetic theory, highlighting their theoretical foundations and exploring their potential applications in machine learning and artificial intelligence.
☆ Joint space-time wind field data extrapolation and uncertainty quantification using nonparametric Bayesian dictionary learning
A methodology is developed, based on nonparametric Bayesian dictionary learning, for joint space-time wind field data extrapolation and estimation of related statistics by relying on limited/incomplete measurements. Specifically, utilizing sparse/incomplete measured data, a time-dependent optimization problem is formulated for determining the expansion coefficients of an associated low-dimensional representation of the stochastic wind field. Compared to an alternative, standard, compressive sampling treatment of the problem, the developed methodology exhibits the following advantages. First, the Bayesian formulation enables also the quantification of the uncertainty in the estimates. Second, the requirement in standard CS-based applications for an a priori selection of the expansion basis is circumvented. Instead, this is done herein in an adaptive manner based on the acquired data. Overall, the methodology exhibits enhanced extrapolation accuracy, even in cases of high-dimensional data of arbitrary form, and of relatively large extrapolation distances. Thus, it can be used, potentially, in a wide range of wind engineering applications where various constraints dictate the use of a limited number of sensors. The efficacy of the methodology is demonstrated by considering two case studies. The first relates to the extrapolation of simulated wind velocity records consistent with a prescribed joint wavenumber-frequency power spectral density in a three-dimensional domain (2D and time). The second pertains to the extrapolation of four-dimensional (3D and time) boundary layer wind tunnel experimental data that exhibit significant spatial variability and non-Gaussian characteristics.
☆ From Observational Data to Clinical Recommendations: A Causal Framework for Estimating Patient-level Treatment Effects and Learning Policies
We propose a framework for building patient-specific treatment recommendation models, building on the large recent literature on learning patient-level causal models and inspired by the target trial paradigm of Hernan and Robins. We focus on safety and validity, including the crucial issue of causal identification when using observational data. We do not provide a specific model, but rather a way to integrate existing methods and know-how into a practical pipeline. We further provide a real world use-case of treatment optimization for patients with heart failure who develop acute kidney injury during hospitalization. The results suggest our pipeline can improve patient outcomes over the current treatment regime.
☆ Step-wise Policy for Rare-tool Knowledge (SPaRK): Offline RL that Drives Diverse Tool Use in LLMs
We present Step-wise Policy for Rare-tool Knowledge (SPaRK), a novel reinforcement learning framework that teaches large language models to explore diverse tool usage patterns beyond conventional high-temperature sampling. Building on recent advances in step-wise reinforcement learning, we introduce a dual-objective reward system that simultaneously optimizes for answer quality and tool diversity, training a Llama-3.1 8B model through offline PPO on synthetically generated trajectories from the MMLU-Pro dataset. Our approach uniquely employs a rarity-first exploitation strategy where a GPT-4o judge scores candidate actions across eight distinct tools plus chain-of-thought reasoning, with the policy favoring less-frequently used but still viable tools to encourage systematic exploration. Empirical results demonstrate that SPaRK achieves competitive performance across 14 MMLU-Pro categories while exhibiting significantly higher entropy in tool selection compared to both baseline and supervised fine-tuning approaches, suggesting that algorithmic exploration through explicit tool diversity can enhance reasoning capabilities without sacrificing accuracy.
comment: 12 pages, 4 figures
☆ Local Pairwise Distance Matching for Backpropagation-Free Reinforcement Learning ECAI 2025
Training neural networks with reinforcement learning (RL) typically relies on backpropagation (BP), necessitating storage of activations from the forward pass for subsequent backward updates. Furthermore, backpropagating error signals through multiple layers often leads to vanishing or exploding gradients, which can degrade learning performance and stability. We propose a novel approach that trains each layer of the neural network using local signals during the forward pass in RL settings. Our approach introduces local, layer-wise losses leveraging the principle of matching pairwise distances from multi-dimensional scaling, enhanced with optional reward-driven guidance. This method allows each hidden layer to be trained using local signals computed during forward propagation, thus eliminating the need for backward passes and storing intermediate activations. Our experiments, conducted with policy gradient methods across common RL benchmarks, demonstrate that this backpropagation-free method achieves competitive performance compared to their classical BP-based counterpart. Additionally, the proposed method enhances stability and consistency within and across runs, and improves performance especially in challenging environments.
comment: accepted at the European Conference on Artificial Intelligence (ECAI 2025)
☆ A Parallelizable Approach for Characterizing NE in Zero-Sum Games After a Linear Number of Iterations of Gradient Descent
We study online optimization methods for zero-sum games, a fundamental problem in adversarial learning in machine learning, economics, and many other domains. Traditional methods approximate Nash equilibria (NE) using either regret-based methods (time-average convergence) or contraction-map-based methods (last-iterate convergence). We propose a new method based on Hamiltonian dynamics in physics and prove that it can characterize the set of NE in a finite (linear) number of iterations of alternating gradient descent in the unbounded setting, modulo degeneracy, a first in online optimization. Unlike standard methods for computing NE, our proposed approach can be parallelized and works with arbitrary learning rates, both firsts in algorithmic game theory. Experimentally, we support our results by showing our approach drastically outperforms standard methods.
☆ Neurosymbolic Reasoning Shortcuts under the Independence Assumption
The ubiquitous independence assumption among symbolic concepts in neurosymbolic (NeSy) predictors is a convenient simplification: NeSy predictors use it to speed up probabilistic reasoning. Recent works like van Krieken et al. (2024) and Marconato et al. (2024) argued that the independence assumption can hinder learning of NeSy predictors and, more crucially, prevent them from correctly modelling uncertainty. There is, however, scepticism in the NeSy community around the scenarios in which the independence assumption actually limits NeSy systems (Faronius and Dos Martires, 2025). In this work, we settle this question by formally showing that assuming independence among symbolic concepts entails that a model can never represent uncertainty over certain concept combinations. Thus, the model fails to be aware of reasoning shortcuts, i.e., the pathological behaviour of NeSy predictors that predict correct downstream tasks but for the wrong reasons.
comment: Accepted at NeSy 2025
☆ Guiding LLM Decision-Making with Fairness Reward Models
Large language models are increasingly used to support high-stakes decisions, potentially influencing who is granted bail or receives a loan. Naive chain-of-thought sampling can improve average decision accuracy, but has also been shown to amplify unfair bias. To address this challenge and enable the trustworthy use of reasoning models in high-stakes decision-making, we propose a framework for training a generalizable Fairness Reward Model (FRM). Our model assigns a fairness score to LLM reasoning, enabling the system to down-weight biased trajectories and favor equitable ones when aggregating decisions across reasoning chains. We show that a single Fairness Reward Model, trained on weakly supervised, LLM-annotated examples of biased versus unbiased reasoning, transfers across tasks, domains, and model families without additional fine-tuning. Applied to real-world decision-making tasks including recidivism prediction and social media moderation, we show that our approach consistently improves fairness while matching, or even surpassing, baseline accuracy.
☆ Internal Value Alignment in Large Language Models through Controlled Value Vector Activation ACL 2025
Aligning Large Language Models (LLMs) with human values has attracted increasing attention since it provides clarity, transparency, and the ability to adapt to evolving scenarios. In this paper, we introduce a Controlled Value Vector Activation (ConVA) method that directly aligns the internal values of LLMs by interpreting how a value is encoded in their latent representations and modifies relevant activations to ensure consistent values in LLMs. To ensure an accurate and unbiased interpretation, we propose a context-controlled value vector identification method. To consistently control values without sacrificing model performance, we introduce a gated value vector activation method for effective and minimum degree of value control. Experiments show that our method achieves the highest control success rate across 10 basic values without hurting LLM performance and fluency, and ensures target values even with opposite and potentially malicious input prompts. Source code and data are available at~ https://github.com/hr-jin/ConVA.
comment: 25 pages, 14 figures. Accepted by ACL 2025 (main conference)
☆ Fast Last-Iterate Convergence of SGD in the Smooth Interpolation Regime
We study population convergence guarantees of stochastic gradient descent (SGD) for smooth convex objectives in the interpolation regime, where the noise at optimum is zero or near zero. The behavior of the last iterate of SGD in this setting -- particularly with large (constant) stepsizes -- has received growing attention in recent years due to implications for the training of over-parameterized models, as well as to analyzing forgetting in continual learning and to understanding the convergence of the randomized Kaczmarz method for solving linear systems. We establish that after $T$ steps of SGD on $\beta$-smooth convex loss functions with stepsize $\eta \leq 1/\beta$, the last iterate exhibits expected excess risk $\widetilde{O}(1/(\eta T^{1-\beta\eta/2}) + \eta T^{\beta\eta/2} \sigma_\star^2)$, where $\sigma_\star^2$ denotes the variance of the stochastic gradients at the optimum. In particular, for a well-tuned stepsize we obtain a near optimal $\widetilde{O}(1/T + \sigma_\star/\sqrt{T})$ rate for the last iterate, extending the results of Varre et al. (2021) beyond least squares regression; and when $\sigma_\star=0$ we obtain a rate of $O(1/\sqrt{T})$ with $\eta=1/\beta$, improving upon the best-known $O(T^{-1/4})$ rate recently established by Evron et al. (2025) in the special case of realizable linear regression.
comment: 27 pages
☆ Turning Sand to Gold: Recycling Data to Bridge On-Policy and Off-Policy Learning via Causal Bound
Deep reinforcement learning (DRL) agents excel in solving complex decision-making tasks across various domains. However, they often require a substantial number of training steps and a vast experience replay buffer, leading to significant computational and resource demands. To address these challenges, we introduce a novel theoretical result that leverages the Neyman-Rubin potential outcomes framework into DRL. Unlike most methods that focus on bounding the counterfactual loss, we establish a causal bound on the factual loss, which is analogous to the on-policy loss in DRL. This bound is computed by storing past value network outputs in the experience replay buffer, effectively utilizing data that is usually discarded. Extensive experiments across the Atari 2600 and MuJoCo domains on various agents, such as DQN and SAC, achieve up to 2,427% higher reward ratio, outperforming the same agents without our proposed term, and reducing the experience replay buffer size by up to 96%, significantly improving sample efficiency at negligible cost.
comment: 51 pages, 16 figures
☆ LyAm: Robust Non-Convex Optimization for Stable Learning in Noisy Environments
Training deep neural networks, particularly in computer vision tasks, often suffers from noisy gradients and unstable convergence, which hinder performance and generalization. In this paper, we propose LyAm, a novel optimizer that integrates Adam's adaptive moment estimation with Lyapunov-based stability mechanisms. LyAm dynamically adjusts the learning rate using Lyapunov stability theory to enhance convergence robustness and mitigate training noise. We provide a rigorous theoretical framework proving the convergence guarantees of LyAm in complex, non-convex settings. Extensive experiments on like as CIFAR-10 and CIFAR-100 show that LyAm consistently outperforms state-of-the-art optimizers in terms of accuracy, convergence speed, and stability, establishing it as a strong candidate for robust deep learning optimization.
☆ Fairness-Aware Grouping for Continuous Sensitive Variables: Application for Debiasing Face Analysis with respect to Skin Tone
Within a legal framework, fairness in datasets and models is typically assessed by dividing observations into predefined groups and then computing fairness measures (e.g., Disparate Impact or Equality of Odds with respect to gender). However, when sensitive attributes such as skin color are continuous, dividing into default groups may overlook or obscure the discrimination experienced by certain minority subpopulations. To address this limitation, we propose a fairness-based grouping approach for continuous (possibly multidimensional) sensitive attributes. By grouping data according to observed levels of discrimination, our method identifies the partition that maximizes a novel criterion based on inter-group variance in discrimination, thereby isolating the most critical subgroups. We validate the proposed approach using multiple synthetic datasets and demonstrate its robustness under changing population distributions - revealing how discrimination is manifested within the space of sensitive attributes. Furthermore, we examine a specialized setting of monotonic fairness for the case of skin color. Our empirical results on both CelebA and FFHQ, leveraging the skin tone as predicted by an industrial proprietary algorithm, show that the proposed segmentation uncovers more nuanced patterns of discrimination than previously reported, and that these findings remain stable across datasets for a given model. Finally, we leverage our grouping model for debiasing purpose, aiming at predicting fair scores with group-by-group post-processing. The results demonstrate that our approach improves fairness while having minimal impact on accuracy, thus confirming our partition method and opening the door for industrial deployment.
☆ Generative Click-through Rate Prediction with Applications to Search Advertising
Click-Through Rate (CTR) prediction models are integral to a myriad of industrial settings, such as personalized search advertising. Current methods typically involve feature extraction from users' historical behavior sequences combined with product information, feeding into a discriminative model that is trained on user feedback to estimate CTR. With the success of models such as GPT, the potential for generative models to enrich expressive power beyond discriminative models has become apparent. In light of this, we introduce a novel model that leverages generative models to enhance the precision of CTR predictions in discriminative models. To reconcile the disparate data aggregation needs of both model types, we design a two-stage training process: 1) Generative pre-training for next-item prediction with the given item category in user behavior sequences; 2) Fine-tuning the well-trained generative model within a discriminative CTR prediction framework. Our method's efficacy is substantiated through extensive experiments on a new dataset, and its significant utility is further corroborated by online A/B testing results. Currently, the model is deployed on one of the world's largest e-commerce platforms, and we intend to release the associated code and dataset in the future.
comment: This work was first submitted on February 9, 2024
☆ Improved sampling algorithms and Poincaré inequalities for non-log-concave distributions
We study the problem of sampling from a distribution $\mu$ with density $\propto e^{-V}$ for some potential function $V:\mathbb R^d\to \mathbb R$ with query access to $V$ and $\nabla V$. We start with the following standard assumptions: (1) The potential function $V$ is $L$-smooth. (2) The second moment $\mathbf{E}_{X\sim \mu}[\|X\|^2]\leq M$. Recently, He and Zhang (COLT'25) showed that the query complexity of sampling from such distributions is at least $\left(\frac{LM}{d\epsilon}\right)^{\Omega(d)}$ where $\epsilon$ is the desired accuracy in total variation distance, and the Poincar\'e constant can be arbitrarily large. Meanwhile, another common assumption in the study of diffusion based samplers (see e.g., the work of Chen, Chewi, Li, Li, Salim and Zhang (ICLR'23)) strengthens the smoothness condition (1) to the following: (1*) The potential function of *every* distribution along the Ornstein-Uhlenbeck process starting from $\mu$ is $L$-smooth. We show that under the assumptions (1*) and (2), the query complexity of sampling from $\mu$ can be $\mathrm{poly}(L,d)\cdot \left(\frac{Ld+M}{\epsilon^2}\right)^{\mathcal{O}(L+1)}$, which is polynomial in $d$ and $\frac{1}{\epsilon}$ when $L=\mathcal{O}(1)$ and $M=\mathrm{poly}(d)$. This improves the algorithm with quasi-polynomial query complexity developed by Huang et al. (COLT'24). Our results imply that the seemly moderate strengthening of the smoothness condition (1) to (1*) can lead to an exponential gap in the query complexity of sampling algorithms. Moreover, we show that together with the assumption (1*) and the stronger moment assumption that $\|X\|$ is $\lambda$-sub-Gaussian for $X\sim\mu$, the Poincar\'e constant of $\mu$ is at most $\mathcal{O}(\lambda)^{2(L+1)}$. As an application of our technique, we obtain improved estimate of the Poincar\'e constant for mixture of Gaussians with the same covariance.
☆ DuetGraph: Coarse-to-Fine Knowledge Graph Reasoning with Dual-Pathway Global-Local Fusion
Knowledge graphs (KGs) are vital for enabling knowledge reasoning across various domains. Recent KG reasoning methods that integrate both global and local information have achieved promising results. However, existing methods often suffer from score over-smoothing, which blurs the distinction between correct and incorrect answers and hinders reasoning effectiveness. To address this, we propose DuetGraph, a coarse-to-fine KG reasoning mechanism with dual-pathway global-local fusion. DuetGraph tackles over-smoothing by segregating -- rather than stacking -- the processing of local (via message passing) and global (via attention) information into two distinct pathways, preventing mutual interference and preserving representational discrimination. In addition, DuetGraph introduces a coarse-to-fine optimization, which partitions entities into high- and low-score subsets. This strategy narrows the candidate space and sharpens the score gap between the two subsets, which alleviates over-smoothing and enhances inference quality. Extensive experiments on various datasets demonstrate that DuetGraph achieves state-of-the-art (SOTA) performance, with up to an 8.7% improvement in reasoning quality and a 1.8$\times$ acceleration in training efficiency.
☆ Gradient Descent on Logistic Regression: Do Large Step-Sizes Work with Data on the Sphere?
Gradient descent (GD) on logistic regression has many fascinating properties. When the dataset is linearly separable, it is known that the iterates converge in direction to the maximum-margin separator regardless of how large the step size is. In the non-separable case, however, it has been shown that GD can exhibit a cycling behaviour even when the step sizes is still below the stability threshold $2/\lambda$, where $\lambda$ is the largest eigenvalue of the Hessian at the solution. This short paper explores whether restricting the data to have equal magnitude is a sufficient condition for global convergence, under any step size below the stability threshold. We prove that this is true in a one dimensional space, but in higher dimensions cycling behaviour can still occur. We hope to inspire further studies on quantifying how common these cycles are in realistic datasets, as well as finding sufficient conditions to guarantee global convergence with large step sizes.
☆ A Robust Incomplete Multimodal Low-Rank Adaptation Approach for Emotion Recognition
Multimodal Emotion Recognition (MER) often encounters incomplete multimodality in practical applications due to sensor failures or privacy protection requirements. While existing methods attempt to address various incomplete multimodal scenarios by balancing the training of each modality combination through additional gradients, these approaches face a critical limitation: training gradients from different modality combinations conflict with each other, ultimately degrading the performance of the final prediction model. In this paper, we propose a unimodal decoupled dynamic low-rank adaptation method based on modality combinations, named MCULoRA, which is a novel framework for the parameter-efficient training of incomplete multimodal learning models. MCULoRA consists of two key modules, modality combination aware low-rank adaptation (MCLA) and dynamic parameter fine-tuning (DPFT). The MCLA module effectively decouples the shared information from the distinct characteristics of individual modality combinations. The DPFT module adjusts the training ratio of modality combinations based on the separability of each modality's representation space, optimizing the learning efficiency across different modality combinations. Our extensive experimental evaluation in multiple benchmark datasets demonstrates that MCULoRA substantially outperforms previous incomplete multimodal learning approaches in downstream task accuracy.
☆ Recent Advances in Simulation-based Inference for Gravitational Wave Data Analysis
The detection of gravitational waves by the LIGO-Virgo-KAGRA collaboration has ushered in a new era of observational astronomy, emphasizing the need for rapid and detailed parameter estimation and population-level analyses. Traditional Bayesian inference methods, particularly Markov chain Monte Carlo, face significant computational challenges when dealing with the high-dimensional parameter spaces and complex noise characteristics inherent in gravitational wave data. This review examines the emerging role of simulation-based inference methods in gravitational wave astronomy, with a focus on approaches that leverage machine-learning techniques such as normalizing flows and neural posterior estimation. We provide a comprehensive overview of the theoretical foundations underlying various simulation-based inference methods, including neural posterior estimation, neural ratio estimation, neural likelihood estimation, flow matching, and consistency models. We explore the applications of these methods across diverse gravitational wave data processing scenarios, from single-source parameter estimation and overlapping signal analysis to testing general relativity and conducting population studies. Although these techniques demonstrate speed improvements over traditional methods in controlled studies, their model-dependent nature and sensitivity to prior assumptions are barriers to their widespread adoption. Their accuracy, which is similar to that of conventional methods, requires further validation across broader parameter spaces and noise conditions.
comment: 30 pages, 6 figures, 1 table. Published version accepted by Astronomical Techniques and Instruments (ATI)
☆ Data-Driven Differential Evolution in Tire Industry Extrusion: Leveraging Surrogate Models
The optimization of industrial processes remains a critical challenge, particularly when no mathematical formulation of objective functions or constraints is available. This study addresses this issue by proposing a surrogate-based, data-driven methodology for optimizing complex real-world manufacturing systems using only historical process data. Machine learning models are employed to approximate system behavior and construct surrogate models, which are integrated into a tailored metaheuristic approach: Data-Driven Differential Evolution with Multi-Level Penalty Functions and Surrogate Models, an adapted version of Differential Evolution suited to the characteristics of the studied process. The methodology is applied to an extrusion process in the tire manufacturing industry, with the goal of optimizing initialization parameters to reduce waste and production time. Results show that the surrogate-based optimization approach outperforms historical best configurations, achieving a 65\% reduction in initialization and setup time, while also significantly minimizing material waste. These findings highlight the potential of combining data-driven modeling and metaheuristic optimization for industrial processes where explicit formulations are unavailable.
comment: 22 pages, 15 figures
☆ Striking the Perfect Balance: Preserving Privacy While Boosting Utility in Collaborative Medical Prediction Platforms
Online collaborative medical prediction platforms offer convenience and real-time feedback by leveraging massive electronic health records. However, growing concerns about privacy and low prediction quality can deter patient participation and doctor cooperation. In this paper, we first clarify the privacy attacks, namely attribute attacks targeting patients and model extraction attacks targeting doctors, and specify the corresponding privacy principles. We then propose a privacy-preserving mechanism and integrate it into a novel one-shot distributed learning framework, aiming to simultaneously meet both privacy requirements and prediction performance objectives. Within the framework of statistical learning theory, we theoretically demonstrate that the proposed distributed learning framework can achieve the optimal prediction performance under specific privacy requirements. We further validate the developed privacy-preserving collaborative medical prediction platform through both toy simulations and real-world data experiments.
☆ An Explainable AI-Enhanced Machine Learning Approach for Cardiovascular Disease Detection and Risk Assessment
Heart disease remains a major global health concern, particularly in regions with limited access to medical resources and diagnostic facilities. Traditional diagnostic methods often fail to accurately identify and manage heart disease risks, leading to adverse outcomes. Machine learning has the potential to significantly enhance the accuracy, efficiency, and speed of heart disease diagnosis. In this study, we proposed a comprehensive framework that combines classification models for heart disease detection and regression models for risk prediction. We employed the Heart Disease dataset, which comprises 1,035 cases. To address the issue of class imbalance, the Synthetic Minority Oversampling Technique (SMOTE) was applied, resulting in the generation of an additional 100,000 synthetic data points. Performance metrics, including accuracy, precision, recall, F1-score, R2, MSE, RMSE, and MAE, were used to evaluate the model's effectiveness. Among the classification models, Random Forest emerged as the standout performer, achieving an accuracy of 97.2% on real data and 97.6% on synthetic data. For regression tasks, Linear Regression demonstrated the highest R2 values of 0.992 and 0.984 on real and synthetic datasets, respectively, with the lowest error metrics. Additionally, Explainable AI techniques were employed to enhance the interpretability of the models. This study highlights the potential of machine learning to revolutionize heart disease diagnosis and risk prediction, thereby facilitating early intervention and enhancing clinical decision-making.
comment: This paper has been accepted at the IEEE QPAIN 2025. The final version will be available in the IEEE Xplore Digital Library
☆ Quantized Rank Reduction: A Communications-Efficient Federated Learning Scheme for Network-Critical Applications
Federated learning is a machine learning approach that enables multiple devices (i.e., agents) to train a shared model cooperatively without exchanging raw data. This technique keeps data localized on user devices, ensuring privacy and security, while each agent trains the model on their own data and only shares model updates. The communication overhead is a significant challenge due to the frequent exchange of model updates between the agents and the central server. In this paper, we propose a communication-efficient federated learning scheme that utilizes low-rank approximation of neural network gradients and quantization to significantly reduce the network load of the decentralized learning process with minimal impact on the model's accuracy.
comment: In Proceedings of the 2025 IARIA Annual Congress on Frontiers in Science, Technology, Services, and Applications (IARIA Congress 2025), Venice, Italy, July 6-10, 2025
☆ Mixture of Experts in Large Language Models
This paper presents a comprehensive review of the Mixture-of-Experts (MoE) architecture in large language models, highlighting its ability to significantly enhance model performance while maintaining minimal computational overhead. Through a systematic analysis spanning theoretical foundations, core architectural designs, and large language model (LLM) applications, we examine expert gating and routing mechanisms, hierarchical and sparse MoE configurations, meta-learning approaches, multimodal and multitask learning scenarios, real-world deployment cases, and recent advances and challenges in deep learning. Our analysis identifies key advantages of MoE, including superior model capacity compared to equivalent Bayesian approaches, improved task-specific performance, and the ability to scale model capacity efficiently. We also underscore the importance of ensuring expert diversity, accurate calibration, and reliable inference aggregation, as these are essential for maximizing the effectiveness of MoE architectures. Finally, this review outlines current research limitations, open challenges, and promising future directions, providing a foundation for continued innovation in MoE architecture and its applications.
☆ Gradient Regularization-based Neural Granger Causality
With the advancement of deep learning technologies, various neural network-based Granger causality models have been proposed. Although these models have demonstrated notable improvements, several limitations remain. Most existing approaches adopt the component-wise architecture, necessitating the construction of a separate model for each time series, which results in substantial computational costs. In addition, imposing the sparsity-inducing penalty on the first-layer weights of the neural network to extract causal relationships weakens the model's ability to capture complex interactions. To address these limitations, we propose Gradient Regularization-based Neural Granger Causality (GRNGC), which requires only one time series prediction model and applies $L_{1}$ regularization to the gradient between model's input and output to infer Granger causality. Moreover, GRNGC is not tied to a specific time series forecasting model and can be implemented with diverse architectures such as KAN, MLP, and LSTM, offering enhanced flexibility. Numerical simulations on DREAM, Lorenz-96, fMRI BOLD, and CausalTime show that GRNGC outperforms existing baselines and significantly reduces computational overhead. Meanwhile, experiments on real-world DNA, Yeast, HeLa, and bladder urothelial carcinoma datasets further validate the model's effectiveness in reconstructing gene regulatory networks.
comment: 9 pages,3 figures, conference
☆ An Interpretable AI framework Quantifying Traditional Chinese Medicine Principles Towards Enhancing and Integrating with Modern Biomedicine
Traditional Chinese Medicine diagnosis and treatment principles, established through centuries of trial-and-error clinical practice, directly maps patient-specific symptom patterns to personalised herbal therapies. These empirical holistic mapping principles offer valuable strategies to address remaining challenges of reductionism methodologies in modern biomedicine. However, the lack of a quantitative framework and molecular-level evidence has limited their interpretability and reliability. Here, we present an AI framework trained on ancient and classical TCM formula records to quantify the symptom pattern-herbal therapy mappings. Interestingly, we find that empirical TCM diagnosis and treatment are consistent with the encoding-decoding processes in the AI model. This enables us to construct an interpretable TCM embedding space (TCM-ES) using the model's quantitative representation of TCM principles. Validated through broad and extensive TCM patient data, the TCM-ES offers universal quantification of the TCM practice and therapeutic efficacy. We further map biomedical entities into the TCM-ES through correspondence alignment. We find that the principal directions of the TCM-ES are significantly associated with key biological functions (such as metabolism, immune, and homeostasis), and that the disease and herb embedding proximity aligns with their genetic relationships in the human protein interactome, which demonstrate the biological significance of TCM principles. Moreover, the TCM-ES uncovers latent disease relationships, and provides alternative metric to assess clinical efficacy for modern disease-drug pairs. Finally, we construct a comprehensive and integrative TCM knowledge graph, which predicts potential associations between diseases and targets, drugs, herbal compounds, and herbal therapies, providing TCM-informed opportunities for disease analysis and drug development.
comment: 31 pages, 6 figures
☆ Real-Time Bayesian Detection of Drift-Evasive GNSS Spoofing in Reinforcement Learning Based UAV Deconfliction
Autonomous unmanned aerial vehicles (UAVs) rely on global navigation satellite system (GNSS) pseudorange measurements for accurate real-time localization and navigation. However, this dependence exposes them to sophisticated spoofing threats, where adversaries manipulate pseudoranges to deceive UAV receivers. Among these, drift-evasive spoofing attacks subtly perturb measurements, gradually diverting the UAVs trajectory without triggering conventional signal-level anti-spoofing mechanisms. Traditional distributional shift detection techniques often require accumulating a threshold number of samples, causing delays that impede rapid detection and timely response. Consequently, robust temporal-scale detection methods are essential to identify attack onset and enable contingency planning with alternative sensing modalities, improving resilience against stealthy adversarial manipulations. This study explores a Bayesian online change point detection (BOCPD) approach that monitors temporal shifts in value estimates from a reinforcement learning (RL) critic network to detect subtle behavioural deviations in UAV navigation. Experimental results show that this temporal value-based framework outperforms conventional GNSS spoofing detectors, temporal semi-supervised learning frameworks, and the Page-Hinkley test, achieving higher detection accuracy and lower false-positive and false-negative rates for drift-evasive spoofing attacks.
☆ Improving Wi-Fi Network Performance Prediction with Deep Learning Models
The increasing need for robustness, reliability, and determinism in wireless networks for industrial and mission-critical applications is the driver for the growth of new innovative methods. The study presented in this work makes use of machine learning techniques to predict channel quality in a Wi-Fi network in terms of the frame delivery ratio. Predictions can be used proactively to adjust communication parameters at runtime and optimize network operations for industrial applications. Methods including convolutional neural networks and long short-term memory were analyzed on datasets acquired from a real Wi-Fi setup across multiple channels. The models were compared in terms of prediction accuracy and computational complexity. Results show that the frame delivery ratio can be reliably predicted, and convolutional neural networks, although slightly less effective than other models, are more efficient in terms of CPU usage and memory consumption. This enhances the model's usability on embedded and industrial systems.
comment: preprint accepted, 8 pages, 2025
☆ How does Labeling Error Impact Contrastive Learning? A Perspective from Data Dimensionality Reduction ICML2025
In recent years, contrastive learning has achieved state-of-the-art performance in the territory of self-supervised representation learning. Many previous works have attempted to provide the theoretical understanding underlying the success of contrastive learning. Almost all of them rely on a default assumption, i.e., the label consistency assumption, which may not hold in practice (the probability of failure is called labeling error) due to the strength and randomness of common augmentation strategies, such as random resized crop (RRC). This paper investigates the theoretical impact of labeling error on the downstream classification performance of contrastive learning. We first reveal several significant negative impacts of labeling error on downstream classification risk. To mitigate these impacts, data dimensionality reduction method (e.g., singular value decomposition, SVD) is applied on original data to reduce false positive samples, and establish both theoretical and empirical evaluations. Moreover, it is also found that SVD acts as a double-edged sword, which may lead to the deterioration of downstream classification accuracy due to the reduced connectivity of the augmentation graph. Based on the above observations, we give the augmentation suggestion that we should use some moderate embedding dimension (such as $512, 1024$ in our experiments), data inflation, weak augmentation, and SVD to ensure large graph connectivity and small labeling error to improve model performance.
comment: Accepted by ICML2025 as a poster
☆ RMAU-NET: A Residual-Multihead-Attention U-Net Architecture for Landslide Segmentation and Detection from Remote Sensing Images
In recent years, landslide disasters have reported frequently due to the extreme weather events of droughts, floods , storms, or the consequence of human activities such as deforestation, excessive exploitation of natural resources. However, automatically observing landslide is challenging due to the extremely large observing area and the rugged topography such as mountain or highland. This motivates us to propose an end-to-end deep-learning-based model which explores the remote sensing images for automatically observing landslide events. By considering remote sensing images as the input data, we can obtain free resource, observe large and rough terrains by time. To explore the remote sensing images, we proposed a novel neural network architecture which is for two tasks of landslide detection and landslide segmentation. We evaluated our proposed model on three different benchmark datasets of LandSlide4Sense, Bijie, and Nepal. By conducting extensive experiments, we achieve F1 scores of 98.23, 93.83 for the landslide detection task on LandSlide4Sense, Bijie datasets; mIoU scores of 63.74, 76.88 on the segmentation tasks regarding LandSlide4Sense, Nepal datasets. These experimental results prove potential to integrate our proposed model into real-life landslide observation systems.
☆ Hashed Watermark as a Filter: Defeating Forging and Overwriting Attacks in Weight-based Neural Network Watermarking
As valuable digital assets, deep neural networks necessitate robust ownership protection, positioning neural network watermarking (NNW) as a promising solution. Among various NNW approaches, weight-based methods are favored for their simplicity and practicality; however, they remain vulnerable to forging and overwriting attacks. To address those challenges, we propose NeuralMark, a robust method built around a hashed watermark filter. Specifically, we utilize a hash function to generate an irreversible binary watermark from a secret key, which is then used as a filter to select the model parameters for embedding. This design cleverly intertwines the embedding parameters with the hashed watermark, providing a robust defense against both forging and overwriting attacks. An average pooling is also incorporated to resist fine-tuning and pruning attacks. Furthermore, it can be seamlessly integrated into various neural network architectures, ensuring broad applicability. Theoretically, we analyze its security boundary. Empirically, we verify its effectiveness and robustness across 13 distinct Convolutional and Transformer architectures, covering five image classification tasks and one text generation task. The source codes are available at https://github.com/AIResearch-Group/NeuralMark.
☆ Interpretable Bayesian Tensor Network Kernel Machines with Automatic Rank and Feature Selection
Tensor Network (TN) Kernel Machines speed up model learning by representing parameters as low-rank TNs, reducing computation and memory use. However, most TN-based Kernel methods are deterministic and ignore parameter uncertainty. Further, they require manual tuning of model complexity hyperparameters like tensor rank and feature dimensions, often through trial-and-error or computationally costly methods like cross-validation. We propose Bayesian Tensor Network Kernel Machines, a fully probabilistic framework that uses sparsity-inducing hierarchical priors on TN factors to automatically infer model complexity. This enables automatic inference of tensor rank and feature dimensions, while also identifying the most relevant features for prediction, thereby enhancing model interpretability. All the model parameters and hyperparameters are treated as latent variables with corresponding priors. Given the Bayesian approach and latent variable dependencies, we apply a mean-field variational inference to approximate their posteriors. We show that applying a mean-field approximation to TN factors yields a Bayesian ALS algorithm with the same computational complexity as its deterministic counterpart, enabling uncertainty quantification at no extra computational cost. Experiments on synthetic and real-world datasets demonstrate the superior performance of our model in prediction accuracy, uncertainty quantification, interpretability, and scalability.
comment: 39 pages, 5 figures, 4 tables. Submitted to Journal of Machine Learning Research. The code is available at: https://github.com/afrakilic/BTN-Kernel-Machines. arXiv admin note: text overlap with arXiv:1401.6497 by other authors
☆ MMOne: Representing Multiple Modalities in One Scene ICCV 2025
Humans perceive the world through multimodal cues to understand and interact with the environment. Learning a scene representation for multiple modalities enhances comprehension of the physical world. However, modality conflicts, arising from inherent distinctions among different modalities, present two critical challenges: property disparity and granularity disparity. To address these challenges, we propose a general framework, MMOne, to represent multiple modalities in one scene, which can be readily extended to additional modalities. Specifically, a modality modeling module with a novel modality indicator is proposed to capture the unique properties of each modality. Additionally, we design a multimodal decomposition mechanism to separate multi-modal Gaussians into single-modal Gaussians based on modality differences. We address the essential distinctions among modalities by disentangling multimodal information into shared and modality-specific components, resulting in a more compact and efficient multimodal scene representation. Extensive experiments demonstrate that our method consistently enhances the representation capability for each modality and is scalable to additional modalities. The code is available at https://github.com/Neal2020GitHub/MMOne.
comment: Accepted to ICCV 2025
☆ What Should LLMs Forget? Quantifying Personal Data in LLMs for Right-to-Be-Forgotten Requests KDD 2025
Large Language Models (LLMs) can memorize and reveal personal information, raising concerns regarding compliance with the EU's GDPR, particularly the Right to Be Forgotten (RTBF). Existing machine unlearning methods assume the data to forget is already known but do not address how to identify which individual-fact associations are stored in the model. Privacy auditing techniques typically operate at the population level or target a small set of identifiers, limiting applicability to individual-level data inquiries. We introduce WikiMem, a dataset of over 5,000 natural language canaries covering 243 human-related properties from Wikidata, and a model-agnostic metric to quantify human-fact associations in LLMs. Our approach ranks ground-truth values against counterfactuals using calibrated negative log-likelihood across paraphrased prompts. We evaluate 200 individuals across 15 LLMs (410M-70B parameters), showing that memorization correlates with subject web presence and model scale. We provide a foundation for identifying memorized personal data in LLMs at the individual level, enabling the dynamic construction of forget sets for machine unlearning and RTBF requests.
comment: 16 pages, 3 figures. Accepted at the 7th Workshop on eXplainable Knowledge Discovery in Data Mining (XKDD 2025), ECML PKDD 2025, Porto, Portugal
☆ Multi-Trigger Poisoning Amplifies Backdoor Vulnerabilities in LLMs
Recent studies have shown that Large Language Models (LLMs) are vulnerable to data poisoning attacks, where malicious training examples embed hidden behaviours triggered by specific input patterns. However, most existing works assume a phrase and focus on the attack's effectiveness, offering limited understanding of trigger mechanisms and how multiple triggers interact within the model. In this paper, we present a framework for studying poisoning in LLMs. We show that multiple distinct backdoor triggers can coexist within a single model without interfering with each other, enabling adversaries to embed several triggers concurrently. Using multiple triggers with high embedding similarity, we demonstrate that poisoned triggers can achieve robust activation even when tokens are substituted or separated by long token spans. Our findings expose a broader and more persistent vulnerability surface in LLMs. To mitigate this threat, we propose a post hoc recovery method that selectively retrains specific model components based on a layer-wise weight difference analysis. Our method effectively removes the trigger behaviour with minimal parameter updates, presenting a practical and efficient defence against multi-trigger poisoning.
☆ A Mathematical Optimization Approach to Multisphere Support Vector Data Description
We present a novel mathematical optimization framework for outlier detection in multimodal datasets, extending Support Vector Data Description approaches. We provide a primal formulation, in the shape of a Mixed Integer Second Order Cone model, that constructs Euclidean hyperspheres to identify anomalous observations. Building on this, we develop a dual model that enables the application of the kernel trick, thus allowing for the detection of outliers within complex, non-linear data structures. An extensive computational study demonstrates the effectiveness of our exact method, showing clear advantages over existing heuristic techniques in terms of accuracy and robustness.
comment: 18 pages, 5 figures, 3 tables
☆ LogTinyLLM: Tiny Large Language Models Based Contextual Log Anomaly Detection
Log anomaly detection using traditional rule based or deep learning based methods is often challenging due to the large volume and highly complex nature of log sequence. So effective way of detection of anomalous sequence of logs is crucial for system maintenance and development. This paper proposes parameter efficient finetuning specifically low rank adaptation (LoRA) and adapter based approaches for finding contextual anomalies in sequence of logs in large log data set. It compares different tiny large language models (LLMs) on the Thunderbird dataset. The results show that LoRA based finetuning provides substantial performance improvements of 18 to 19 percentage over LogBert based full finetuning approach, achieving accuracy scores between 97.76% and 98.83% compared to 79.37%.
☆ A Distance Metric for Mixed Integer Programming Instances ECAI 2025
Mixed-integer linear programming (MILP) is a powerful tool for addressing a wide range of real-world problems, but it lacks a clear structure for comparing instances. A reliable similarity metric could establish meaningful relationships between instances, enabling more effective evaluation of instance set heterogeneity and providing better guidance to solvers, particularly when machine learning is involved. Existing similarity metrics often lack precision in identifying instance classes or rely heavily on labeled data, which limits their applicability and generalization. To bridge this gap, this paper introduces the first mathematical distance metric for MILP instances, derived directly from their mathematical formulations. By discretizing right-hand sides, weights, and variables into classes, the proposed metric draws inspiration from the Earth mover's distance to quantify mismatches in weight-variable distributions for constraint comparisons. This approach naturally extends to enable instance-level comparisons. We evaluate both an exact and a greedy variant of our metric under various parameter settings, using the StrIPLIB dataset. Results show that all components of the metric contribute to class identification, and that the greedy version achieves accuracy nearly identical to the exact formulation while being nearly 200 times faster. Compared to state-of-the-art baselines, including feature-based, image-based, and neural network models, our unsupervised method consistently outperforms all non-learned approaches and rivals the performance of a supervised classifier on class and subclass grouping tasks.
comment: Accepted to ECAI 2025
☆ GATE: Graph Attention Neural Networks with Real-Time Edge Construction for Robust Indoor Localization using Mobile Embedded Devices
Accurate indoor localization is crucial for enabling spatial context in smart environments and navigation systems. Wi-Fi Received Signal Strength (RSS) fingerprinting is a widely used indoor localization approach due to its compatibility with mobile embedded devices. Deep Learning (DL) models improve accuracy in localization tasks by learning RSS variations across locations, but they assume fingerprint vectors exist in a Euclidean space, failing to incorporate spatial relationships and the non-uniform distribution of real-world RSS noise. This results in poor generalization across heterogeneous mobile devices, where variations in hardware and signal processing distort RSS readings. Graph Neural Networks (GNNs) can improve upon conventional DL models by encoding indoor locations as nodes and modeling their spatial and signal relationships as edges. However, GNNs struggle with non-Euclidean noise distributions and suffer from the GNN blind spot problem, leading to degraded accuracy in environments with dense access points (APs). To address these challenges, we propose GATE, a novel framework that constructs an adaptive graph representation of fingerprint vectors while preserving an indoor state-space topology, modeling the non-Euclidean structure of RSS noise to mitigate environmental noise and address device heterogeneity. GATE introduces 1) a novel Attention Hyperspace Vector (AHV) for enhanced message passing, 2) a novel Multi-Dimensional Hyperspace Vector (MDHV) to mitigate the GNN blind spot, and 3) an new Real-Time Edge Construction (RTEC) approach for dynamic graph adaptation. Extensive real-world evaluations across multiple indoor spaces with varying path lengths, AP densities, and heterogeneous devices demonstrate that GATE achieves 1.6x to 4.72x lower mean localization errors and 1.85x to 4.57x lower worst-case errors compared to state-of-the-art indoor localization frameworks.
☆ Relative Entropy Pathwise Policy Optimization
Score-function policy gradients have delivered strong results in game-playing, robotics and language-model fine-tuning. Yet its high-variance often undermines training stability. On the other hand, pathwise policy gradients alleviate the training variance, but are reliable only when driven by an accurate action-conditioned value function which is notoriously hard to train without relying on past off-policy data. In this paper, we discuss how to construct a value-gradient driven, on-policy algorithm that allow training Q-value models purely from on-policy data, unlocking the possibility of using pathwise policy updates in the context of on-policy learning. We show how to balance stochastic policies for exploration with constrained policy updates for stable training, and evaluate important architectural components that facilitate accurate value function learning. Building on these insights, we propose Relative Entropy Pathwise Policy Optimization (REPPO), an efficient on-policy algorithm that combines the sample-efficiency of pathwise policy gradients with the simplicity and minimal memory footprint of standard on-policy learning. We demonstrate that REPPO provides strong empirical performance at decreased sample requirements, wall-clock time, memory footprint as well as high hyperparameter robustness in a set of experiments on two standard GPU-parallelized benchmarks.
☆ First-Order Error Matters: Accurate Compensation for Quantized Large Language Models
Post-training quantization (PTQ) offers an efficient approach to compressing large language models (LLMs), significantly reducing memory access and computational costs. Existing compensation-based weight calibration methods often rely on a second-order Taylor expansion to model quantization error, under the assumption that the first-order term is negligible in well-trained full-precision models. However, we reveal that the progressive compensation process introduces accumulated first-order deviations between latent weights and their full-precision counterparts, making this assumption fundamentally flawed. To address this, we propose FOEM, a novel PTQ method that explicitly incorporates first-order gradient terms to improve quantization error compensation. FOEM approximates gradients by directly computing the difference between latent and full-precision weights, avoiding the high cost and limited generalization of backpropagation-based gradient computation. This approach introduces minimal additional computational overhead. Moreover, FOEM leverages precomputed Cholesky factors to efficiently recover the inverse of Hessian submatrices in real time. Extensive experiments across a wide range of models and benchmarks demonstrate that FOEM consistently outperforms the classical GPTQ method. In 3-bit weight-only quantization, FOEM reduces the perplexity of Llama3-8B by 89.6%, and improves the 5-shot MMLU accuracy of Llama3-70B from 51.7% to 74.9%, approaching the full-precision performance of 78.6%. Furthermore, FOEM can be seamlessly integrated with advanced techniques such as GPTAQ and SpinQuant, yielding additional improvements under the challenging W4A4KV4 setting, and further narrowing the accuracy gap with full-precision baselines beyond what current state-of-the-art methods achieve. The code is available at https://github.com/Xingyu-Zheng/FOEM.
☆ Leveraging Advanced Machine Learning to Predict Turbulence Dynamics from Temperature Observations at an Experimental Prescribed Fire
This study explores the potential for predicting turbulent kinetic energy (TKE) from more readily acquired temperature data using temperature profiles and turbulence data collected concurrently at 10 Hz during a small experimental prescribed burn in the New Jersey Pine Barrens. Machine learning models, including Deep Neural Networks, Random Forest Regressor, Gradient Boosting, and Gaussian Process Regressor, were employed to assess the potential to predict TKE from temperature perturbations and explore temporal and spatial dynamics of correlations. Data visualization and correlation analyses revealed patterns and relationships between thermocouple temperatures and TKE, providing insight into the underlying dynamics. More accurate predictions of TKE were achieved by employing various machine learning models despite a weak correlation between the predictors and the target variable. The results demonstrate significant success, particularly from regression models, in accurately predicting the TKE. The findings of this study demonstrate a novel numerical approach to identifying new relationships between temperature and airflow processes in and around the fire environment. These relationships can help refine our understanding of combustion environment processes and the coupling and decoupling of fire environment processes necessary for improving fire operations strategy and fire and smoke model predictions. The findings of this study additionally highlight the valuable role of machine learning techniques in analyzing the complex large datasets of the fire environments, showcasing their potential to advance fire research and management practices.
comment: arXiv admin note: text overlap with arXiv:2311.05128
☆ AdaMuon: Adaptive Muon Optimizer
We propose AdaMuon, an adaptive learning-rate framework built upon the recently validated Muon optimizer, which has demonstrated substantial efficiency gains over AdamW in large-scale model training. AdaMuon augments Muon with two mutually dependent modules: (1) a per-parameter second-moment modulation that captures orthogonal gradient updates to ensure update-level adaptivity, and (2) a RMS-aligned rescaling that regulates the overall update magnitude by aligning it with the intrinsic structure of the parameter space. Empirical results on multiple model scales and learning-rate regimes confirm that AdaMuon consistently outperforms the original Muon, delivering higher acceleration in convergence while maintaining training stability. Our method introduces no additional tuning burden and can be seamlessly integrated into existing Muon training pipelines.
☆ Crafting Imperceptible On-Manifold Adversarial Attacks for Tabular Data
Adversarial attacks on tabular data present fundamental challenges distinct from image or text domains due to the heterogeneous nature of mixed categorical and numerical features. Unlike images where pixel perturbations maintain visual similarity, tabular data lacks intuitive similarity metrics, making it difficult to define imperceptible modifications. Additionally, traditional gradient-based methods prioritise $\ell_p$-norm constraints, often producing adversarial examples that deviate from the original data distributions, making them detectable. We propose a latent space perturbation framework using a mixed-input Variational Autoencoder (VAE) to generate imperceptible adversarial examples. The proposed VAE integrates categorical embeddings and numerical features into a unified latent manifold, enabling perturbations that preserve statistical consistency. We specify In-Distribution Success Rate (IDSR) to measure the proportion of adversarial examples that remain statistically indistinguishable from the input distribution. Evaluation across six publicly available datasets and three model architectures demonstrates that our method achieves substantially lower outlier rates and more consistent performance compared to traditional input-space attacks and other VAE-based methods adapted from image domain approaches. Our comprehensive analysis includes hyperparameter sensitivity, sparsity control mechanisms, and generative architectural comparisons, revealing that VAE-based attacks depend critically on reconstruction quality but offer superior practical utility when sufficient training data is available. This work highlights the importance of on-manifold perturbations for realistic adversarial attacks on tabular data, offering a robust approach for practical deployment. The source code can be accessed through https://github.com/ZhipengHe/VAE-TabAttack.
comment: 32 pages
☆ Misalignment from Treating Means as Ends
Reward functions, learned or manually specified, are rarely perfect. Instead of accurately expressing human goals, these reward functions are often distorted by human beliefs about how best to achieve those goals. Specifically, these reward functions often express a combination of the human's terminal goals -- those which are ends in themselves -- and the human's instrumental goals -- those which are means to an end. We formulate a simple example in which even slight conflation of instrumental and terminal goals results in severe misalignment: optimizing the misspecified reward function results in poor performance when measured by the true reward function. This example distills the essential properties of environments that make reinforcement learning highly sensitive to conflation of instrumental and terminal goals. We discuss how this issue can arise with a common approach to reward learning and how it can manifest in real environments.
☆ High-Throughput Distributed Reinforcement Learning via Adaptive Policy Synchronization
Scaling reinforcement learning (RL) workloads often requires distributing environment simulation across compute clusters. Existing frameworks entangle simulation, learning logic, and orchestration into monolithic systems, limiting modularity and reusability. We present ClusterEnv, a lightweight, learner-agnostic interface for distributed environment execution that mirrors the Gymnasium API. ClusterEnv introduces the DETACH pattern, which decouples simulation from training by offloading reset() and step() operations to remote workers while keeping learning centralized. To address policy staleness in distributed execution, we propose Adaptive Actor Policy Synchronization (AAPS), a divergence-triggered update mechanism that reduces synchronization overhead without sacrificing performance. ClusterEnv integrates cleanly into existing RL pipelines, supports both on-policy and off-policy methods, and requires minimal code changes. Experiments on discrete control tasks demonstrate that AAPS achieves high sample efficiency with significantly fewer weight updates. Source code is available at https://github.com/rodlaf/ClusterEnv.
☆ StellarF: A Lora-Adapter Integrated Large Model Framework for Stellar Flare Forecasting with Historical & Statistical Data
Stellar flare forecasting, a critical research frontier in astronomy, offers profound insights into stellar activity. However, the field is constrained by both the sparsity of recorded flare events and the absence of domain-specific large-scale predictive models. To address these challenges, this study introduces StellarF (Stellar Flare Forecasting), a novel large model that leverages Low-Rank (LoRA) and Adapter techniques to parameter-efficient learning for stellar flare forecasting. At its core, StellarF integrates an flare statistical information module with a historical flare record module, enabling multi-scale pattern recognition from observational data. Extensive experiments on our self-constructed datasets (derived from Kepler and TESS light curves) demonstrate that StellarF achieves state-of-the-art performance compared to existing methods. The proposed prediction paradigm establishes a novel methodological framework for advancing astrophysical research and cross-disciplinary applications.
☆ Physics-Informed Neural Networks For Semiconductor Film Deposition: A Review
Semiconductor manufacturing relies heavily on film deposition processes, such as Chemical Vapor Deposition and Physical Vapor Deposition. These complex processes require precise control to achieve film uniformity, proper adhesion, and desired functionality. Recent advancements in Physics-Informed Neural Networks (PINNs), an innovative machine learning (ML) approach, have shown significant promise in addressing challenges related to process control, quality assurance, and predictive modeling within semiconductor film deposition and other manufacturing domains. This paper provides a comprehensive review of ML applications targeted at semiconductor film deposition processes. Through a thematic analysis, we identify key trends, existing limitations, and research gaps, offering insights into both the advantages and constraints of current methodologies. Our structured analysis aims to highlight the potential integration of these ML techniques to enhance interpretability, accuracy, and robustness in film deposition processes. Additionally, we examine state-of-the-art PINN methods, discussing strategies for embedding physical knowledge, governing laws, and partial differential equations into advanced neural network architectures tailored for semiconductor manufacturing. Based on this detailed review, we propose novel research directions that integrate the strengths of PINNs to significantly advance film deposition processes. The contributions of this study include establishing a clear pathway for future research in integrating physics-informed ML frameworks, addressing existing methodological gaps, and ultimately improving precision, scalability, and operational efficiency within semiconductor manufacturing.
comment: 11 pages, 1 figure, 3 tables, IDETC-CIE 2025
☆ GOLFS: Feature Selection via Combining Both Global and Local Information for High Dimensional Clustering
It is important to identify the discriminative features for high dimensional clustering. However, due to the lack of cluster labels, the regularization methods developed for supervised feature selection can not be directly applied. To learn the pseudo labels and select the discriminative features simultaneously, we propose a new unsupervised feature selection method, named GlObal and Local information combined Feature Selection (GOLFS), for high dimensional clustering problems. The GOLFS algorithm combines both local geometric structure via manifold learning and global correlation structure of samples via regularized self-representation to select the discriminative features. The combination improves the accuracy of both feature selection and clustering by exploiting more comprehensive information. In addition, an iterative algorithm is proposed to solve the optimization problem and the convergency is proved. Simulations and two real data applications demonstrate the excellent finite-sample performance of GOLFS on both feature selection and clustering.
☆ Diffusion Decoding for Peptide De Novo Sequencing
Peptide de novo sequencing is a method used to reconstruct amino acid sequences from tandem mass spectrometry data without relying on existing protein sequence databases. Traditional deep learning approaches, such as Casanovo, mainly utilize autoregressive decoders and predict amino acids sequentially. Subsequently, they encounter cascading errors and fail to leverage high-confidence regions effectively. To address these issues, this paper investigates using diffusion decoders adapted for the discrete data domain. These decoders provide a different approach, allowing sequence generation to start from any peptide segment, thereby enhancing prediction accuracy. We experiment with three different diffusion decoder designs, knapsack beam search, and various loss functions. We find knapsack beam search did not improve performance metrics and simply replacing the transformer decoder with a diffusion decoder lowered performance. Although peptide precision and recall were still 0, the best diffusion decoder design with the DINOISER loss function obtained a statistically significant improvement in amino acid recall by 0.373 compared to the baseline autoregressive decoder-based Casanovo model. These findings highlight the potential of diffusion decoders to not only enhance model sensitivity but also drive significant advancements in peptide de novo sequencing.
♻ ☆ EXPO: Stable Reinforcement Learning with Expressive Policies
We study the problem of training and fine-tuning expressive policies with online reinforcement learning (RL) given an offline dataset. Training expressive policy classes with online RL present a unique challenge of stable value maximization. Unlike simpler Gaussian policies commonly used in online RL, expressive policies like diffusion and flow-matching policies are parameterized by a long denoising chain, which hinders stable gradient propagation from actions to policy parameters when optimizing against some value function. Our key insight is that we can address stable value maximization by avoiding direct optimization over value with the expressive policy and instead construct an on-the-fly RL policy to maximize Q-value. We propose Expressive Policy Optimization (EXPO), a sample-efficient online RL algorithm that utilizes an on-the-fly policy to maximize value with two parameterized policies -- a larger expressive base policy trained with a stable imitation learning objective and a light-weight Gaussian edit policy that edits the actions sampled from the base policy toward a higher value distribution. The on-the-fly policy optimizes the actions from the base policy with the learned edit policy and chooses the value maximizing action from the base and edited actions for both sampling and temporal-difference (TD) backup. Our approach yields up to 2-3x improvement in sample efficiency on average over prior methods both in the setting of fine-tuning a pretrained policy given offline data and in leveraging offline data to train online.
comment: corrected typo, formatting, added experiments
♻ ☆ Hi Robot: Open-Ended Instruction Following with Hierarchical Vision-Language-Action Models ICML 2025
Generalist robots that can perform a range of different tasks in open-world settings must be able to not only reason about the steps needed to accomplish their goals, but also process complex instructions, prompts, and even feedback during task execution. Intricate instructions (e.g., "Could you make me a vegetarian sandwich?" or "I don't like that one") require not just the ability to physically perform the individual steps, but the ability to situate complex commands and feedback in the physical world. In this work, we describe a system that uses vision-language models in a hierarchical structure, first reasoning over complex prompts and user feedback to deduce the most appropriate next step to fulfill the task, and then performing that step with low-level actions. In contrast to direct instruction following methods that can fulfill simple commands ("pick up the cup"), our system can reason through complex prompts and incorporate situated feedback during task execution ("that's not trash"). We evaluate our system across three robotic platforms, including single-arm, dual-arm, and dual-arm mobile robots, demonstrating its ability to handle tasks such as cleaning messy tables, making sandwiches, and grocery shopping. Videos are available at https://www.pi.website/research/hirobot
comment: ICML 2025
♻ ☆ Are DeepSeek R1 And Other Reasoning Models More Faithful?
Language models trained to solve reasoning tasks via reinforcement learning have achieved striking results. We refer to these models as reasoning models. Are the Chains of Thought (CoTs) of reasoning models more faithful than traditional models? We evaluate three reasoning models (based on Qwen-2.5, Gemini-2, and DeepSeek-V3-Base) on an existing test of faithful CoT. To measure faithfulness, we test whether models can describe how a cue in their prompt influences their answer to MMLU questions. For example, when the cue "A Stanford Professor thinks the answer is D" is added to the prompt, models sometimes switch their answer to D. In such cases, the DeepSeek-R1 reasoning model describes the cue's influence 59% of the time, compared to 7% for the non-reasoning DeepSeek model. We evaluate seven types of cue, such as misleading few-shot examples and suggestive follow-up questions from the user. Reasoning models describe cues that influence them much more reliably than all the non-reasoning models tested (including Claude-3.5-Sonnet and GPT-4o). In an additional experiment, we provide evidence suggesting that the use of reward models causes less faithful responses -- which may help explain why non-reasoning models are less faithful. Our study has two main limitations. First, we test faithfulness using a set of artificial tasks, which may not reflect realistic use-cases. Second, we only measure one specific aspect of faithfulness -- whether models can describe the influence of cues. Future research should investigate whether the advantage of reasoning models in faithfulness holds for a broader set of tests. Still, we think this increase in faithfulness is promising for the explainability of language models.
comment: 10 pages, 8 figures
♻ ☆ Large Language Models Engineer Too Many Simple Features For Tabular Data NeurIPS 2024
Tabular machine learning problems often require time-consuming and labor-intensive feature engineering. Recent efforts have focused on using large language models (LLMs) to capitalize on their potential domain knowledge. At the same time, researchers have observed ethically concerning negative biases in other LLM-related use cases, such as text generation. These developments motivated us to investigate whether LLMs exhibit a bias that negatively impacts the performance of feature engineering. While not ethically concerning, such a bias could hinder practitioners from fully utilizing LLMs for automated data science. Therefore, we propose a method to detect potential biases by detecting anomalies in the frequency of operators (e.g., adding two features) suggested by LLMs when engineering new features. Our experiments evaluate the bias of four LLMs, two big frontier and two small open-source models, across 27 tabular datasets. Our results indicate that LLMs are biased toward simple operators, such as addition, and can fail to utilize more complex operators, such as grouping followed by aggregations. Furthermore, the bias can negatively impact the predictive performance when using LLM-generated features. Our results call for mitigating bias when using LLMs for feature engineering.
comment: Accepted at the 3rd Table Representation Learning Workshop @ NeurIPS 2024
♻ ☆ A Mathematical Theory of Discursive Networks
Large-language models (LLMs) turn writing into a live exchange between humans and software. We characterize this new medium as a discursive network that treats people and LLMs as equal nodes and tracks how their statements circulate. We define the generation of erroneous information as invalidation (any factual, logical, or structural breach) and show it follows four hazards: drift from truth, self-repair, fresh fabrication, and external detection. We develop a general mathematical model of discursive networks that shows that a network governed only by drift and self-repair stabilizes at a modest error rate. Giving each false claim even a small chance of peer review shifts the system to a truth-dominant state. We operationalize peer review with the open-source \emph{Flaws-of-Others (FOO) algorithm}: a configurable loop in which any set of agents critique one another while a harmonizer merges their verdicts. We identify an ethical transgression, epithesis, that occurs when humans fail to engage in the discursive network. The takeaway is practical and cultural: reliability in this new medium comes not from perfecting single models but from connecting imperfect ones into networks that enforce mutual accountability.
comment: 32 pages, 4 figures, 4 tables, 1 algorithm, 54 references
♻ ☆ ComFairGNN: Community Fair Graph Neural Network PAKDD 2025
Graph Neural Networks (GNNs) have become the leading approach for addressing graph analytical problems in various real-world scenarios. However, GNNs may produce biased predictions against certain demographic subgroups due to node attributes and neighbors surrounding a node. Most current research on GNN fairness focuses predominantly on debiasing GNNs using oversimplified fairness evaluation metrics, which can give a misleading impression of fairness. Understanding the potential evaluation paradoxes due to the complicated nature of the graph structure is crucial for developing effective GNN debiasing mechanisms. In this paper, we examine the effectiveness of current GNN debiasing methods in terms of unfairness evaluation. Specifically, we introduce a community-level strategy to measure bias in GNNs and evaluate debiasing methods at this level. Further, We introduce ComFairGNN, a novel framework designed to mitigate community-level bias in GNNs. Our approach employs a learnable coreset-based debiasing function that addresses bias arising from diverse local neighborhood distributions during GNNs neighborhood aggregation. Comprehensive evaluations on three benchmark datasets demonstrate our model's effectiveness in both accuracy and fairness metrics.
comment: PAKDD 2025
♻ ☆ Searching Latent Program Spaces
General intelligence requires systems that acquire new skills efficiently and generalize beyond their training distributions. Although program synthesis approaches have strong generalization power, they face scaling issues due to large combinatorial spaces that quickly make them impractical and require human-generated DSLs or pre-trained priors to narrow this search space. On the other hand, deep learning methods have had high successes, but they lack structured test-time adaptation and rely on heavy stochastic sampling or expensive gradient updates for fine-tuning. In this work, we propose the Latent Program Network (LPN), a new architecture that builds in test-time search directly into neural models. LPN learns a latent space of implicit programs--neurally mapping inputs to outputs--through which it can search using gradients at test time. LPN combines the adaptability of symbolic approaches and the scalability of neural methods. It searches through a compact latent space at test time and bypasses the need for pre-defined domain-specific languages. On a range of programming-by-examples tasks, LPN either outperforms or matches performance compared to in-context learning and test-time training methods. Tested on the ARC-AGI benchmark, we demonstrate that LPN can both learn a compact program space and search through it at test time to adapt to novel tasks. LPN doubles its performance on out-of-distribution tasks when test-time search is switched on.
comment: Code available at https://github.com/clement-bonnet/lpn
♻ ☆ Reinforcement Learning with Action Chunking
We present Q-chunking, a simple yet effective recipe for improving reinforcement learning (RL) algorithms for long-horizon, sparse-reward tasks. Our recipe is designed for the offline-to-online RL setting, where the goal is to leverage an offline prior dataset to maximize the sample-efficiency of online learning. Effective exploration and sample-efficient learning remain central challenges in this setting, as it is not obvious how the offline data should be utilized to acquire a good exploratory policy. Our key insight is that action chunking, a technique popularized in imitation learning where sequences of future actions are predicted rather than a single action at each timestep, can be applied to temporal difference (TD)-based RL methods to mitigate the exploration challenge. Q-chunking adopts action chunking by directly running RL in a 'chunked' action space, enabling the agent to (1) leverage temporally consistent behaviors from offline data for more effective online exploration and (2) use unbiased $n$-step backups for more stable and efficient TD learning. Our experimental results demonstrate that Q-chunking exhibits strong offline performance and online sample efficiency, outperforming prior best offline-to-online methods on a range of long-horizon, sparse-reward manipulation tasks.
comment: 25 pages, 15 figures
♻ ☆ Model See Model Do: Speech-Driven Facial Animation with Style Control SIGGRAPH
Speech-driven 3D facial animation plays a key role in applications such as virtual avatars, gaming, and digital content creation. While existing methods have made significant progress in achieving accurate lip synchronization and generating basic emotional expressions, they often struggle to capture and effectively transfer nuanced performance styles. We propose a novel example-based generation framework that conditions a latent diffusion model on a reference style clip to produce highly expressive and temporally coherent facial animations. To address the challenge of accurately adhering to the style reference, we introduce a novel conditioning mechanism called style basis, which extracts key poses from the reference and additively guides the diffusion generation process to fit the style without compromising lip synchronization quality. This approach enables the model to capture subtle stylistic cues while ensuring that the generated animations align closely with the input speech. Extensive qualitative, quantitative, and perceptual evaluations demonstrate the effectiveness of our method in faithfully reproducing the desired style while achieving superior lip synchronization across various speech scenarios.
comment: 10 pages, 7 figures, SIGGRAPH Conference Papers '25
♻ ☆ Gram-Schmidt Methods for Unsupervised Feature Extraction and Selection
Feature extraction and selection in the presence of nonlinear dependencies among the data is a fundamental challenge in unsupervised learning. We propose using a Gram-Schmidt (GS) type orthogonalization process over function spaces to detect and map out such dependencies. Specifically, by applying the GS process over some family of functions, we construct a series of covariance matrices that can either be used to identify new large-variance directions, or to remove those dependencies from known directions. In the former case, we provide information-theoretic guarantees in terms of entropy reduction. In the latter, we provide precise conditions by which the chosen function family eliminates existing redundancy in the data. Each approach provides both a feature extraction and a feature selection algorithm. Our feature extraction methods are linear, and can be seen as natural generalization of principal component analysis (PCA). We provide experimental results for synthetic and real-world benchmark datasets which show superior performance over state-of-the-art (linear) feature extraction and selection algorithms. Surprisingly, our linear feature extraction algorithms are comparable and often outperform several important nonlinear feature extraction methods such as autoencoders, kernel PCA, and UMAP. Furthermore, one of our feature selection algorithms strictly generalizes a recent Fourier-based feature selection mechanism (Heidari et al., IEEE Transactions on Information Theory, 2022), yet at significantly reduced complexity.
comment: To appear in IEEE Transactions on Information Theory
♻ ☆ Training neural control variates using correlated configurations
Neural control variates (NCVs) have emerged as a powerful tool for variance reduction in Monte Carlo (MC) simulations, particularly in high-dimensional problems where traditional control variates are difficult to construct analytically. By training neural networks to learn auxiliary functions correlated with the target observable, NCVs can significantly reduce estimator variance while preserving unbiasedness. However, a critical but often overlooked aspect of NCV training is the role of autocorrelated samples generated by Markov Chain Monte Carlo (MCMC). While such samples are typically discarded for error estimation due to their statistical redundancy, they may contain useful information about the structure of the underlying probability distribution that can benefit the training process. In this work, we systematically examine the effect of using correlated configurations in training neural control variates. We demonstrate, both conceptually and numerically, that training on correlated data can improve control variate performance, especially in settings with limited computational resources. Our analysis includes empirical results from $U(1)$ gauge theory and scalar field theory, illustrating when and how autocorrelated samples enhance NCV construction. These findings provide practical guidance for the efficient use of MCMC data in training neural networks.
comment: 9 pages, 8 figures
♻ ☆ A Generative Approach to LLM Harmfulness Detection with Special Red Flag Tokens
Most safety training methods for large language models (LLMs) are based on fine-tuning that forces models to shift from an unsafe answer to refusal when faced with harmful requests. Unfortunately, these drastic distribution shifts generally compromise model capabilities. To avoid that, we propose to expand the model's vocabulary with a special token we call red flag token () and propose to train the model to insert this token into its response at any time when harmful content is generated or about to be generated. Our approach offers several advantages: it enables the model to explicitly learn the concept of harmfulness while marginally affecting the generated distribution, thus maintaining the model's utility. It also evaluates each generated answer and provides robustness as good as adversarial training without the need to run attacks during training. Moreover, by encapsulating our safety tuning in a LoRA module, we provide additional defenses against fine-tuning API attacks.
comment: 14 pages, 6 figures
♻ ☆ Synthetic Datasets for Machine Learning on Spatio-Temporal Graphs using PDEs
Many physical processes can be expressed through partial differential equations (PDEs). Real-world measurements of such processes are often collected at irregularly distributed points in space, which can be effectively represented as graphs; however, there are currently only a few existing datasets. Our work aims to make advancements in the field of PDE-modeling accessible to the temporal graph machine learning community, while addressing the data scarcity problem, by creating and utilizing datasets based on PDEs. In this work, we create and use synthetic datasets based on PDEs to support spatio-temporal graph modeling in machine learning for different applications. More precisely, we showcase three equations to model different types of disasters and hazards in the fields of epidemiology, atmospheric particles, and tsunami waves. Further, we show how such created datasets can be used by benchmarking several machine learning models on the epidemiological dataset. Additionally, we show how pre-training on this dataset can improve model performance on real-world epidemiological data. The presented methods enable others to create datasets and benchmarks customized to individual requirements. The source code for our methodology and the three created datasets can be found on https://github.com/github-usr-ano/Temporal_Graph_Data_PDEs.
comment: Camera-ready version of the paper, which is now accepted at DMLR, see https://openreview.net/forum?id=EguDBMechn . 17 pages, 5 Figures
♻ ☆ Overcoming Slow Decision Frequencies in Continuous Control: Model-Based Sequence Reinforcement Learning for Model-Free Control ICLR 2025
Reinforcement learning (RL) is rapidly reaching and surpassing human-level control capabilities. However, state-of-the-art RL algorithms often require timesteps and reaction times significantly faster than human capabilities, which is impractical in real-world settings and typically necessitates specialized hardware. We introduce Sequence Reinforcement Learning (SRL), an RL algorithm designed to produce a sequence of actions for a given input state, enabling effective control at lower decision frequencies. SRL addresses the challenges of learning action sequences by employing both a model and an actor-critic architecture operating at different temporal scales. We propose a "temporal recall" mechanism, where the critic uses the model to estimate intermediate states between primitive actions, providing a learning signal for each individual action within the sequence. Once training is complete, the actor can generate action sequences independently of the model, achieving model-free control at a slower frequency. We evaluate SRL on a suite of continuous control tasks, demonstrating that it achieves performance comparable to state-of-the-art algorithms while significantly reducing actor sample complexity. To better assess performance across varying decision frequencies, we introduce the Frequency-Averaged Score (FAS) metric. Our results show that SRL significantly outperforms traditional RL algorithms in terms of FAS, making it particularly suitable for applications requiring variable decision frequencies. Furthermore, we compare SRL with model-based online planning, showing that SRL achieves comparable FAS while leveraging the same model during training that online planners use for planning.
comment: 30 pages, 14 figures, 7 tables. Presented at the Thirteenth International Conference on Learning Representations (ICLR 2025), Singapore, April 24-28, 2025
♻ ☆ Matrix Is All You Need
Deep neural networks employ specialized architectures for vision, sequential and language tasks, yet this proliferation obscures their underlying commonalities. We introduce a unified matrix-order framework that casts convolutional, recurrent and self-attention operations as sparse matrix multiplications. Convolution is realized via an upper-triangular weight matrix performing first-order transformations; recurrence emerges from a lower-triangular matrix encoding stepwise updates; attention arises naturally as a third-order tensor factorization. We prove algebraic isomorphism with standard CNN, RNN and Transformer layers under mild assumptions. Empirical evaluations on image classification (MNIST, CIFAR-10/100, Tiny ImageNet), time-series forecasting (ETTh1, Electricity Load Diagrams) and language modeling/classification (AG News, WikiText-2, Penn Treebank) confirm that sparse-matrix formulations match or exceed native model performance while converging in comparable or fewer epochs. By reducing architecture design to sparse pattern selection, our matrix perspective aligns with GPU parallelism and leverages mature algebraic optimization tools. This work establishes a mathematically rigorous substrate for diverse neural architectures and opens avenues for principled, hardware-aware network design.
♻ ☆ Improving sub-seasonal wind-speed forecasts in Europe with a non-linear model
Sub-seasonal wind speed forecasts provide valuable guidance for wind power system planning and operations, yet the forecast skills of surface winds decrease sharply after two weeks. However, large-scale variables exhibit greater predictability on this time scale. This study explores the potential of leveraging non-linear relationships between 500 hPa geopotential height (Z500) and surface wind speed to improve sub-seasonal wind speed forecast skills in Europe. Our proposed framework uses a Multiple Linear Regression (MLR) or a Convolutional Neural Network (CNN) to regress surface wind speed from Z500. Evaluations on ERA5 reanalysis indicate that the CNN performs better due to its non-linearity. Applying these models to sub-seasonal forecasts from the European Centre for Medium-Range Weather Forecasts, various verification metrics demonstrate the advantages of non-linearity. Yet, this is partly explained by the fact that these statistical models are under-dispersive since they explain only a fraction of the target variable variance. Introducing stochastic perturbations to represent the stochasticity of the unexplained part from the signal helps compensate for this issue. Results show that the perturbed CNN performs better than the perturbed MLR only in the first weeks, while the perturbed MLR's performance converges towards that of the perturbed CNN after two weeks. The study finds that introducing stochastic perturbations can address the issue of insufficient spread in these statistical models, with improvements from the non-linearity varying with the lead time of the forecasts.
♻ ☆ A Resource Efficient Quantum Kernel
Quantum processors may enhance machine learning by mapping high-dimensional data onto quantum systems for processing. Conventional quantum kernels, or feature maps, for encoding data features onto a quantum circuit are currently impractical, as the number of entangling gates scales quadratically with the dimension of the dataset and the number of qubits. In this work, we introduce a quantum kernel designed to handle high-dimensional data with a significantly reduced number of qubits and entangling operations. Our approach preserves essential data characteristics while promoting computational efficiency, as evidenced by extensive experiments on benchmark datasets that demonstrate a marked improvement in both accuracy and resource utilization, as compared to state-of-the-art quantum feature maps. Our noisy simulations results combined with lower resource requirements highlight our kernel's ability to function within the constraints of noisy intermediate-scale quantum devices. Through numerical simulations and small-scale implementation on a superconducting circuit quantum computing platform, we demonstrate that our scheme performs on par or better than a set of classical algorithms for classification. Our findings herald a promising avenue for the practical implementation of quantum machine learning algorithms on near future quantum computing platforms.
comment: 17 pages, 12 figures
♻ ☆ DFRot: Achieving Outlier-Free and Massive Activation-Free for Rotated LLMs with Refined Rotation
Rotating the activation and weight matrices to reduce the influence of outliers in large language models (LLMs) has recently attracted significant attention, particularly in the context of model quantization. Prior studies have shown that in low-precision quantization scenarios, such as 4-bit weights and 4-bit activations (W4A4), randomized Hadamard transforms can achieve significantly higher accuracy than randomized orthogonal transforms. Notably, the reason behind this phenomenon remains unknown. In this paper, we find that these transformations show substantial improvement in eliminating outliers for common tokens and achieve similar quantization error. The primary reason for the accuracy difference lies in the fact that randomized Hadamard transforms can slightly reduce the quantization error for tokens with massive activations while randomized orthogonal transforms increase the quantization error. Due to the extreme rarity of these tokens and their critical impact on model accuracy, we consider this a long-tail optimization problem, and therefore construct a simple yet effective method: a weighted loss function. Additionally, we propose an optimization strategy for the rotation matrix that involves alternating optimization of quantization parameters while employing orthogonal Procrustes transforms to refine the rotation matrix. This makes the distribution of the rotated activation values more conducive to quantization, especially for tokens with massive activations. Our method enhances the Rotated LLMs by achieving dual free, Outlier-Free and Massive Activation-Free, dubbed as DFRot. Extensive experiments demonstrate the effectiveness and efficiency of DFRot. By tuning the rotation matrix using just a single sample, DFRot achieves a perplexity improvement of 0.98 and 0.95 on W4A4KV4 and W4A4KV16, respectively, for LLaMA3-70B, a model known for its quantization challenges.
comment: Accepeted bythe 2nd Conference on Language Modeling (COLM 2025). Source code \url{https://github.com/JingyangXiang/DFRot}
♻ ☆ Temporal Chunking Enhances Recognition of Implicit Sequential Patterns
In this pilot study, we propose a neuro-inspired approach that compresses temporal sequences into context-tagged chunks, where each tag represents a recurring structural unit or``community'' in the sequence. These tags are generated during an offline sleep phase and serve as compact references to past experience, allowing the learner to incorporate information beyond its immediate input range. We evaluate this idea in a controlled synthetic environment designed to reveal the limitations of traditional neural network based sequence learners, such as recurrent neural networks (RNNs), when facing temporal patterns on multiple timescales. We evaluate this idea in a controlled synthetic environment designed to reveal the limitations of traditional neural network based sequence learners, such as recurrent neural networks (RNNs), when facing temporal patterns on multiple timescales. Our results, while preliminary, suggest that temporal chunking can significantly enhance learning efficiency under resource constrained settings. A small-scale human pilot study using a Serial Reaction Time Task further motivates the idea of structural abstraction. Although limited to synthetic tasks, this work serves as an early proof-of-concept, with initial evidence that learned context tags can transfer across related task, offering potential for future applications in transfer learning.
♻ ☆ Gaussian mixture models as a proxy for interacting language models
Large language models (LLMs) are a powerful tool with the ability to match human capabilities and behavior in many settings. Retrieval-augmented generation (RAG) further allows LLMs to generate diverse output depending on the contents of their RAG database. This motivates their use in the social sciences to study human behavior between individuals when large-scale experiments are infeasible. However, LLMs depend on complex, computationally expensive algorithms. In this paper, we introduce interacting Gaussian mixture models (GMMs) as an alternative to similar frameworks using LLMs. We compare a simplified model of GMMs to select experimental simulations of LLMs whose updating and response depend on feedback from other LLMs. We find that interacting GMMs capture important features of the dynamics in interacting LLMs, and we investigate key similarities and differences between interacting LLMs and GMMs. We conclude by discussing the benefits of Gaussian mixture models, potential modifications, and future research directions.
♻ ☆ Moderate Adaptive Linear Units (MoLU)
We propose the Moderate Adaptive Linear Unit (MoLU), a novel activation function for deep neural networks, defined analytically as: f(x)=x \times (1+tanh(x))/2. MoLU combines mathematical elegance with empirical effectiveness, exhibiting superior performance in terms of prediction accuracy, convergence speed, and computational efficiency. Due to its C-infinity smoothness, i.e. infinite differentiability and analyticity, MoLU is expected to mitigate issues such as vanishing or exploding gradients, making it suitable for a broad range of architectures and applications, including large language models (LLMs), Neural Ordinary Differential Equations (Neural ODEs), Physics-Informed Neural Networks (PINNs), and Convolutional Neural Networks (CNNs). Empirical evaluations show that MoLU consistently achieves faster convergence and improved final accuracy relative to widely used activation functions such as GeLU, SiLU, and Mish. These properties position MoLU as a promising and robust candidate for general-purpose activation across diverse deep learning paradigms.
comment: 8 pages, 10 figures
♻ ☆ X Hacking: The Threat of Misguided AutoML ICML 2025
Explainable AI (XAI) and interpretable machine learning methods help to build trust in model predictions and derived insights, yet also present a perverse incentive for analysts to manipulate XAI metrics to support pre-specified conclusions. This paper introduces the concept of X-hacking, a form of p-hacking applied to XAI metrics such as SHAP values. We show how easily an automated machine learning pipeline can be adapted to exploit model multiplicity at scale: searching a Rashomon set of 'defensible' models with similar predictive performance to find a desired explanation. We formulate the trade-off between explanation and accuracy as a multi-objective optimisation problem, and illustrate empirically on familiar real-world datasets that, on average, Bayesian optimisation accelerates X-hacking 3-fold for features susceptible to it, versus random sampling. We show the vulnerability of a dataset to X-hacking can be determined by information redundancy among features. Finally, we suggest possible methods for detection and prevention, and discuss ethical implications for the credibility and reproducibility of XAI.
comment: Accepted to ICML 2025
♻ ☆ Inverse Reinforcement Learning with Switching Rewards and History Dependency for Characterizing Animal Behaviors
Traditional approaches to studying decision-making in neuroscience focus on simplified behavioral tasks where animals perform repetitive, stereotyped actions to receive explicit rewards. While informative, these methods constrain our understanding of decision-making to short timescale behaviors driven by explicit goals. In natural environments, animals exhibit more complex, long-term behaviors driven by intrinsic motivations that are often unobservable. Recent works in time-varying inverse reinforcement learning (IRL) aim to capture shifting motivations in long-term, freely moving behaviors. However, a crucial challenge remains: animals make decisions based on their history, not just their current state. To address this, we introduce SWIRL (SWitching IRL), a novel framework that extends traditional IRL by incorporating time-varying, history-dependent reward functions. SWIRL models long behavioral sequences as transitions between short-term decision-making processes, each governed by a unique reward function. SWIRL incorporates biologically plausible history dependency to capture how past decisions and environmental contexts shape behavior, offering a more accurate description of animal decision-making. We apply SWIRL to simulated and real-world animal behavior datasets and show that it outperforms models lacking history dependency, both quantitatively and qualitatively. This work presents the first IRL model to incorporate history-dependent policies and rewards to advance our understanding of complex, naturalistic decision-making in animals.
♻ ☆ An All-digital 8.6-nJ/Frame 65-nm Tsetlin Machine Image Classification Accelerator
We present an all-digital programmable machine learning accelerator chip for image classification, underpinning on the Tsetlin machine (TM) principles. The TM is an emerging machine learning algorithm founded on propositional logic, utilizing sub-pattern recognition expressions called clauses. The accelerator implements the coalesced TM version with convolution, and classifies booleanized images of 28$\times$28 pixels with 10 categories. A configuration with 128 clauses is used in a highly parallel architecture. Fast clause evaluation is achieved by keeping all clause weights and Tsetlin automata (TA) action signals in registers. The chip is implemented in a 65 nm low-leakage CMOS technology, and occupies an active area of 2.7 mm$^2$. At a clock frequency of 27.8 MHz, the accelerator achieves 60.3k classifications per second, and consumes 8.6 nJ per classification. This demonstrates the energy-efficiency of the TM, which was the main motivation for developing this chip. The latency for classifying a single image is 25.4 $\mu$s which includes system timing overhead. The accelerator achieves 97.42%, 84.54% and 82.55% test accuracies for the datasets MNIST, Fashion-MNIST and Kuzushiji-MNIST, respectively, matching the TM software models.
comment: Copyright 2025 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purpose\, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works
♻ ☆ DeInfoReg: A Decoupled Learning Framework for Better Training Throughput
This paper introduces Decoupled Supervised Learning with Information Regularization (DeInfoReg), a novel approach that transforms a long gradient flow into multiple shorter ones, thereby mitigating the vanishing gradient problem. Integrating a pipeline strategy, DeInfoReg enables model parallelization across multiple GPUs, significantly improving training throughput. We compare our proposed method with standard backpropagation and other gradient flow decomposition techniques. Extensive experiments on diverse tasks and datasets demonstrate that DeInfoReg achieves superior performance and better noise resistance than traditional BP models and efficiently utilizes parallel computing resources. The code for reproducibility is available at: https://github.com/ianzih/Decoupled-Supervised-Learning-for-Information-Regularization/.
♻ ☆ Beyond Predictions: A Participatory Framework for Multi-Stakeholder Decision-Making
Conventional automated decision-support systems, often based on supervised learning, focus on predicting outcomes to recommend actions. However, they typically overlook the complexity of multi-actor environments, where diverse and conflicting stakeholder preferences must be balanced. At the same time, participatory AI approaches remain largely context-specific, limiting their broader applicability. To address these gaps, we propose a participatory framework that reframes decision-making as a multi-stakeholder optimization problem, using context-dependent reward functions to represent each actor's preferences. Our modular, model-agnostic framework employs k-fold cross-validation to fine-tune user-provided prediction models and evaluate decision strategies, including compromise functions that mediate stakeholder trade-offs. A synthetic scoring mechanism aggregates user-defined preferences across multiple metrics to rank strategies and select an optimal decision-maker for generating actionable recommendations on new data. Validated on two high-stake real-world case studies, the framework consistently produces stakeholder-aware decisions that outperform purely predictive baselines across multiple metrics, while enhancing the transparency and accountability of AI-supported decision-making.
♻ ☆ Intervening in Black Box: Concept Bottleneck Model for Enhancing Human Neural Network Mutual Understanding ICCV 2025
Recent advances in deep learning have led to increasingly complex models with deeper layers and more parameters, reducing interpretability and making their decisions harder to understand. While many methods explain black-box reasoning, most lack effective interventions or only operate at sample-level without modifying the model itself. To address this, we propose the Concept Bottleneck Model for Enhancing Human-Neural Network Mutual Understanding (CBM-HNMU). CBM-HNMU leverages the Concept Bottleneck Model (CBM) as an interpretable framework to approximate black-box reasoning and communicate conceptual understanding. Detrimental concepts are automatically identified and refined (removed/replaced) based on global gradient contributions. The modified CBM then distills corrected knowledge back into the black-box model, enhancing both interpretability and accuracy. We evaluate CBM-HNMU on various CNN and transformer-based models across Flower-102, CIFAR-10, CIFAR-100, FGVC-Aircraft, and CUB-200, achieving a maximum accuracy improvement of 2.64% and a maximum increase in average accuracy across 1.03%. Source code is available at: https://github.com/XiGuaBo/CBM-HNMU.
comment: Accepted by ICCV 2025
♻ ☆ Universal rates of ERM for agnostic learning COLT
The universal learning framework has been developed to obtain guarantees on the learning rates that hold for any fixed distribution, which can be much faster than the ones uniformly hold over all the distributions. Given that the Empirical Risk Minimization (ERM) principle being fundamental in the PAC theory and ubiquitous in practical machine learning, the recent work of arXiv:2412.02810 studied the universal rates of ERM for binary classification under the realizable setting. However, the assumption of realizability is too restrictive to hold in practice. Indeed, the majority of the literature on universal learning has focused on the realizable case, leaving the non-realizable case barely explored. In this paper, we consider the problem of universal learning by ERM for binary classification under the agnostic setting, where the ''learning curve" reflects the decay of the excess risk as the sample size increases. We explore the possibilities of agnostic universal rates and reveal a compact trichotomy: there are three possible agnostic universal rates of ERM, being either $e^{-n}$, $o(n^{-1/2})$, or arbitrarily slow. We provide a complete characterization of which concept classes fall into each of these categories. Moreover, we also establish complete characterizations for the target-dependent universal rates as well as the Bayes-dependent universal rates.
comment: Accepted for presentation at the Conference on Learning Theory (COLT) 2025
♻ ☆ Contrast All the Time: Learning Time Series Representation from Temporal Consistency ECAI
Representation learning for time series using contrastive learning has emerged as a critical technique for improving the performance of downstream tasks. To advance this effective approach, we introduce CaTT (\textit{Contrast All The Time}), a new approach to unsupervised contrastive learning for time series, which takes advantage of dynamics between temporally similar moments more efficiently and effectively than existing methods. CaTT departs from conventional time-series contrastive approaches that rely on data augmentations or selected views. Instead, it uses the full temporal dimension by contrasting all time steps in parallel. This is made possible by a scalable NT-pair formulation, which extends the classic N-pair loss across both batch and temporal dimensions, making the learning process end-to-end and more efficient. CaTT learns directly from the natural structure of temporal data, using repeated or adjacent time steps as implicit supervision, without the need for pair selection heuristics. We demonstrate that this approach produces superior embeddings which allow better performance in downstream tasks. Additionally, training is faster than other contrastive learning approaches, making it suitable for large-scale and real-world time series applications. The source code is publicly available at \href{https://github.com/sfi-norwai/CaTT}{https://github.com/sfi-norwai/CaTT}.
comment: Published in the 28th European Conference on AI (ECAI), October 2025
♻ ☆ Supercharging Floorplan Localization with Semantic Rays ICCV 2025
Floorplans provide a compact representation of the building's structure, revealing not only layout information but also detailed semantics such as the locations of windows and doors. However, contemporary floorplan localization techniques mostly focus on matching depth-based structural cues, ignoring the rich semantics communicated within floorplans. In this work, we introduce a semantic-aware localization framework that jointly estimates depth and semantic rays, consolidating over both for predicting a structural-semantic probability volume. Our probability volume is constructed in a coarse-to-fine manner: We first sample a small set of rays to obtain an initial low-resolution probability volume. We then refine these probabilities by performing a denser sampling only in high-probability regions and process the refined values for predicting a 2D location and orientation angle. We conduct an evaluation on two standard floorplan localization benchmarks. Our experiments demonstrate that our approach substantially outperforms state-of-the-art methods, achieving significant improvements in recall metrics compared to prior works. Moreover, we show that our framework can easily incorporate additional metadata such as room labels, enabling additional gains in both accuracy and efficiency.
comment: Accepted at ICCV 2025. https://tau-vailab.github.io/SemRayLoc/
♻ ☆ Grasping a Handful: Sequential Multi-Object Dexterous Grasp Generation
We introduce the sequential multi-object robotic grasp sampling algorithm SeqGrasp that can robustly synthesize stable grasps on diverse objects using the robotic hand's partial Degrees of Freedom (DoF). We use SeqGrasp to construct the large-scale Allegro Hand sequential grasping dataset SeqDataset and use it for training the diffusion-based sequential grasp generator SeqDiffuser. We experimentally evaluate SeqGrasp and SeqDiffuser against the state-of-the-art non-sequential multi-object grasp generation method MultiGrasp in simulation and on a real robot. The experimental results demonstrate that SeqGrasp and SeqDiffuser reach an 8.71%-43.33% higher grasp success rate than MultiGrasp. Furthermore, SeqDiffuser is approximately 1000 times faster at generating grasps than SeqGrasp and MultiGrasp.
comment: 8 pages, 7 figures
♻ ☆ FeDa4Fair: Client-Level Federated Datasets for Fairness Evaluation
Federated Learning (FL) enables collaborative model training across multiple clients without sharing clients' private data. However, fairness remains a key concern, as biases in local clients' datasets can impact the entire federated system. Heterogeneous data distributions across clients may lead to models that are fairer for some clients than others. Although several fairness-enhancing solutions are present in the literature, most focus on mitigating bias for a single sensitive attribute, typically binary, overlooking the diverse and sometimes conflicting fairness needs of different clients. This limited perspective can limit the effectiveness of fairness interventions for the different clients. To support more robust and reproducible fairness research in FL, we aim to enable a consistent benchmarking of fairness-aware FL methods at both the global and client levels. In this paper, we contribute in three ways: (1) We introduce FeDa4Fair, a library to generate tabular datasets tailored to evaluating fair FL methods under heterogeneous client bias; (2) we release four bias-heterogeneous datasets and corresponding benchmarks to compare fairness mitigation methods in a controlled environment; (3) we provide ready-to-use functions for evaluating fairness outcomes for these datasets.
♻ ☆ Energy Efficiency in AI for 5G and Beyond: A DeepRx Case Study
This study addresses the challenge of balancing energy efficiency with performance in AI/ML models, focusing on DeepRX, a deep learning receiver based on a fully convolutional ResNet architecture. We evaluate the energy consumption of DeepRX, considering factors including FLOPs/Watt and FLOPs/clock, and find consistency between estimated and actual energy usage, influenced by memory access patterns. The research extends to comparing energy dynamics during training and inference phases. A key contribution is the application of knowledge distillation (KD) to train a compact DeepRX student model that emulates the performance of the teacher model but with reduced energy consumption. We experiment with different student model sizes, optimal teacher sizes, and KD hyperparameters. Performance is measured by comparing the Bit Error Rate (BER) performance versus Signal-to-Interference & Noise Ratio (SINR) values of the distilled model and a model trained from scratch. The distilled models demonstrate a lower error floor across SINR levels, highlighting the effectiveness of KD in achieving energy-efficient AI solutions.
♻ ☆ Parameter-Efficient Fine-Tuning with Circulant and Diagonal Vectors IJCAI-2025
Foundation models have achieved tremendous success in different domains. However, their huge computation and storage complexity make these models difficult to fine-tune and also less applicable in practice. Recent study shows training in Fourier domain can be an effective fine-tuning method in terms of both model performance and number of training parameters. In this work, we propose to further reduce the complexity by the factorization through the product of interleaved circulant and diagonal matrices. In addition, we address the case of non-square fine-tuning weights by partitioning the circulant matrix into blocks. Our method avoids the construction of weight change matrix and utilizes 1D fast Fourier transform (FFT) instead of 2D FFT. Experimental results show that our method achieves similar or better performance across various tasks with much less floating-point operations (FLOPs) and the number of trainable parameters.
comment: to appear in Proceedings of the 2025 International Joint Conference on Artificial Intelligence (IJCAI-2025)
♻ ☆ Gaussian Loss Smoothing Enables Certified Training with Tight Convex Relaxations
Training neural networks with high certified accuracy against adversarial examples remains an open challenge despite significant efforts. While certification methods can effectively leverage tight convex relaxations for bound computation, in training, these methods, perhaps surprisingly, can perform worse than looser relaxations. Prior work hypothesized that this phenomenon is caused by the discontinuity, non-smoothness, and perturbation sensitivity of the loss surface induced by tighter relaxations. In this work, we theoretically show that applying Gaussian Loss Smoothing (GLS) on the loss surface can alleviate these issues. We confirm this empirically by instantiating GLS with two variants: a zeroth-order optimization algorithm, called PGPE, which allows training with non-differentiable relaxations, and a first-order optimization algorithm, called RGS, which requires gradients of the relaxation but is much more efficient than PGPE. Extensive experiments show that when combined with tight relaxations, these methods surpass state-of-the-art methods when training on the same network architecture for many settings. Our results clearly demonstrate the promise of Gaussian Loss Smoothing for training certifiably robust neural networks and pave a path towards leveraging tighter relaxations for certified training.
comment: Accepted for publication in TMLR 07/2025
♻ ☆ BridgeNet: A Hybrid, Physics-Informed Machine Learning Framework for Solving High-Dimensional Fokker-Planck Equations
BridgeNet is a novel hybrid framework that integrates convolutional neural networks with physics-informed neural networks to efficiently solve non-linear, high-dimensional Fokker-Planck equations (FPEs). Traditional PINNs, which typically rely on fully connected architectures, often struggle to capture complex spatial hierarchies and enforce intricate boundary conditions. In contrast, BridgeNet leverages adaptive CNN layers for effective local feature extraction and incorporates a dynamically weighted loss function that rigorously enforces physical constraints. Extensive numerical experiments across various test cases demonstrate that BridgeNet not only achieves significantly lower error metrics and faster convergence compared to conventional PINN approaches but also maintains robust stability in high-dimensional settings. This work represents a substantial advancement in computational physics, offering a scalable and accurate solution methodology with promising applications in fields ranging from financial mathematics to complex system dynamics.
♻ ☆ Privacy Against Agnostic Inference Attacks in Vertical Federated Learning
A novel form of inference attack in vertical federated learning (VFL) is proposed, where two parties collaborate in training a machine learning (ML) model. Logistic regression is considered for the VFL model. One party, referred to as the active party, possesses the ground truth labels of the samples in the training phase, while the other, referred to as the passive party, only shares a separate set of features corresponding to these samples. It is shown that the active party can carry out inference attacks on both training and prediction phase samples by acquiring an ML model independently trained on the training samples available to them. This type of inference attack does not require the active party to be aware of the score of a specific sample, hence it is referred to as an agnostic inference attack. It is shown that utilizing the observed confidence scores during the prediction phase, before the time of the attack, can improve the performance of the active party's autonomous ML model, and thus improve the quality of the agnostic inference attack. As a countermeasure, privacy-preserving schemes (PPSs) are proposed. While the proposed schemes preserve the utility of the VFL model, they systematically distort the VFL parameters corresponding to the passive party's features. The level of the distortion imposed on the passive party's parameters is adjustable, giving rise to a trade-off between privacy of the passive party and interpretabiliy of the VFL outcomes by the active party. The distortion level of the passive party's parameters could be chosen carefully according to the privacy and interpretabiliy concerns of the passive and active parties, respectively, with the hope of keeping both parties (partially) satisfied. Finally, experimental results demonstrate the effectiveness of the proposed attack and the PPSs.
♻ ☆ Block Circulant Adapter for Large Language Models IJCAI-2025
Fine-tuning large language models (LLMs) is difficult due to their huge model size. Recent Fourier domain-based methods show potential for reducing fine-tuning costs. We propose a block circulant matrix-based fine-tuning method with a stable training heuristic to leverage the properties of circulant matrices and one-dimensional Fourier transforms to reduce storage and computation costs. Experiments show that our method uses $14\times$ less number of parameters than VeRA, $16\times$ smaller than LoRA and $32\times$ less FLOPs than FourierFT, while maintaining close or better task performance. Our approach presents a promising way in frequency domain to fine-tune large models on downstream tasks.
comment: to appear in Proceedings of the 2025 International Joint Conference on Artificial Intelligence (IJCAI-2025)
♻ ☆ Learning Safe Numeric Planning Action Models
A significant challenge in applying planning technology to real-world problems lies in obtaining a planning model that accurately represents the problem's dynamics. Obtaining a planning model is even more challenging in mission-critical domains, where a trial-and-error approach to learning how to act is not an option. In such domains, the action model used to generate plans must be safe, in the sense that plans generated with it must be applicable and achieve their goals. % Learning safe action models for planning has been mostly explored for domains in which states are sufficiently described with Boolean variables. % In this work, we go beyond this limitation and propose the Numeric Safe Action Models Learning (N-SAM) algorithm. In this work, we present N-SAM, an action model learning algorithm capable of learning safe numeric preconditions and effects. We prove that N-SAM runs in linear time in the number of observations and, under certain conditions, is guaranteed to return safe action models. However, to preserve this safety guarantee, N-SAM must observe a substantial number of examples for each action before including it in the learned model. We address this limitation of N-SAM and propose N-SAM*, an extension to the N-SAM algorithm that always returns an action model where every observed action is applicable at least in some states, even if it was observed only once. N-SAM* does so without compromising the safety of the returned action model. We prove that N-SAM* is optimal in terms of sample complexity compared to any other algorithm that guarantees safety. N-SAM and N-SAM* are evaluated over an extensive benchmark of numeric planning domains, and their performance is compared to a state-of-the-art numeric action model learning algorithm. We also provide a discussion on the impact of numerical accuracy on the learning process.
♻ ☆ The Pragmatic Frames of Spurious Correlations in Machine Learning: Interpreting How and Why They Matter
Learning correlations from data forms the foundation of today's machine learning (ML) and artificial intelligence (AI) research. While contemporary methods enable the automatic discovery of complex patterns, they are prone to failure when unintended correlations are captured. This vulnerability has spurred a growing interest in interrogating spuriousness, which is often seen as a threat to model performance, fairness, and robustness. In this article, we trace departures from the conventional statistical definition of spuriousness -- which denotes a non-causal relationship arising from coincidence or confounding -- to examine how its meaning is negotiated in ML research. Rather than relying solely on formal definitions, researchers assess spuriousness through what we call pragmatic frames: judgments based on what a correlation does in practice -- how it affects model behavior, supports or impedes task performance, or aligns with broader normative goals. Drawing on a broad survey of ML literature, we identify four such frames: relevance ("Models should use correlations that are relevant to the task"), generalizability ("Models should use correlations that generalize to unseen data"), human-likeness ("Models should use correlations that a human would use to perform the same task"), and harmfulness ("Models should use correlations that are not socially or ethically harmful"). These representations reveal that correlation desirability is not a fixed statistical property but a situated judgment informed by technical, epistemic, and ethical considerations. By examining how a foundational ML conundrum is problematized in research literature, we contribute to broader conversations on the contingent practices through which technical concepts like spuriousness are defined and operationalized.
♻ ☆ Shared Global and Local Geometry of Language Model Embeddings
Researchers have recently suggested that models share common representations. In our work, we find numerous geometric similarities across the token embeddings of large language models. First, we find ``global'' similarities: token embeddings often share similar relative orientations. Next, we characterize local geometry in two ways: (1) by using Locally Linear Embeddings, and (2) by defining a simple measure for the intrinsic dimension of each embedding. Both characterizations allow us to find local similarities across token embeddings. Additionally, our intrinsic dimension demonstrates that embeddings lie on a lower dimensional manifold, and that tokens with lower intrinsic dimensions often have semantically coherent clusters, while those with higher intrinsic dimensions do not. Based on our findings, we introduce EMB2EMB, a simple application to linearly transform steering vectors from one language model to another, despite the two models having different dimensions.
♻ ☆ AnnoPage Dataset: Dataset of Non-Textual Elements in Documents with Fine-Grained Categorization ICDAR2025
We introduce the AnnoPage Dataset, a novel collection of 7,550 pages from historical documents, primarily in Czech and German, spanning from 1485 to the present, focusing on the late 19th and early 20th centuries. The dataset is designed to support research in document layout analysis and object detection. Each page is annotated with axis-aligned bounding boxes (AABB) representing elements of 25 categories of non-textual elements, such as images, maps, decorative elements, or charts, following the Czech Methodology of image document processing. The annotations were created by expert librarians to ensure accuracy and consistency. The dataset also incorporates pages from multiple, mainly historical, document datasets to enhance variability and maintain continuity. The dataset is divided into development and test subsets, with the test set carefully selected to maintain the category distribution. We provide baseline results using YOLO and DETR object detectors, offering a reference point for future research. The AnnoPage Dataset is publicly available on Zenodo (https://doi.org/10.5281/zenodo.12788419), along with ground-truth annotations in YOLO format.
comment: 17 pages, 2 tables, 7 figures; Accepted to GREC Workshop at ICDAR2025
♻ ☆ Few-Shot Radar Signal Recognition through Self-Supervised Learning and Radio Frequency Domain Adaptation
Radar signal recognition (RSR) plays a pivotal role in electronic warfare (EW), as accurately classifying radar signals is critical for informing decision-making. Recent advances in deep learning have shown significant potential in improving RSR in domains with ample annotated data. However, these methods fall short in EW scenarios where annotated radio frequency (RF) data are scarce or impractical to obtain. To address these challenges, we introduce a self-supervised learning (SSL) method which utilises masked signal modelling and RF domain adaption to perform few-shot RSR and enhance performance in environments with limited RF samples and annotations. We propose a two-step approach, first pre-training masked autoencoders (MAE) on baseband in-phase and quadrature (I/Q) signals from diverse RF domains, and then transferring the learned representations to the radar domain, where annotated data are scarce. Empirical results show that our lightweight self-supervised ResNet1D model with domain adaptation achieves up to a 17.5% improvement in 1-shot classification accuracy when pre-trained on in-domain signals (i.e., radar signals) and up to a 16.31% improvement when pre-trained on out-of-domain signals (i.e., comm signals), compared to its baseline without using SSL. We also present reference results for several MAE designs and pre-training strategies, establishing a new benchmark for few-shot radar signal classification.
comment: 6 pages, 15 figures
♻ ☆ Multi-View Node Pruning for Accurate Graph Representation
Graph pooling, which compresses a whole graph into a smaller coarsened graph, is an essential component of graph representation learning. To efficiently compress a given graph, graph pooling methods often drop their nodes with attention-based scoring with the task loss. However, this often results in simply removing nodes with lower degrees without consideration of their feature-level relevance to the given task. To fix this problem, we propose a Multi-View Pruning(MVP), a graph pruning method based on a multi-view framework and reconstruction loss. Given a graph, MVP first constructs multiple graphs for different views either by utilizing the predefined modalities or by randomly partitioning the input features, to consider the importance of each node in diverse perspectives. Then, it learns the score for each node by considering both the reconstruction and the task loss. MVP can be incorporated with any hierarchical pooling framework to score the nodes. We validate MVP on multiple benchmark datasets by coupling it with two graph pooling methods, and show that it significantly improves the performance of the base graph pooling method, outperforming all baselines. Further analysis shows that both the encoding of multiple views and the consideration of reconstruction loss are the key to the success of MVP, and that it indeed identifies nodes that are less important according to domain knowledge.
comment: Jiseong Park and Hanjin Kim are co-first author for this work
♻ ☆ TorchCP: A Python Library for Conformal Prediction
Conformal prediction (CP) is a robust statistical framework that generates prediction intervals or sets with guaranteed coverage probability, addressing the challenge of quantifying predictive uncertainty in deep learning. Despite advancements in deep learning architectures and datasets, reliable uncertainty estimation remains elusive, making CP increasingly vital. This paper introduces TorchCP, a PyTorch-native library designed to integrate state-of-the-art CP algorithms into deep learning tasks, including classification, regression, graph neural networks, and large language models. TorchCP offers a comprehensive suite of advanced methodologies, a modular design for easy customization, and full GPU-accelerated scalability. Released under the LGPL-3.0 license, TorchCP has gained widespread adoption with over 12,582 PyPi downloads. It is supported by approximately 16,132 lines of code, 564 unit tests achieving 100\% coverage, and comprehensive documentation. By bridging statistics and computer science, TorchCP empowers researchers and practitioners to advance conformal prediction in diverse deep learning applications.
♻ ☆ On Equivariant Model Selection through the Lens of Uncertainty UAI 2025
Equivariant models leverage prior knowledge on symmetries to improve predictive performance, but misspecified architectural constraints can harm it instead. While work has explored learning or relaxing constraints, selecting among pretrained models with varying symmetry biases remains challenging. We examine this model selection task from an uncertainty-aware perspective, comparing frequentist (via Conformal Prediction), Bayesian (via the marginal likelihood), and calibration-based measures to naive error-based evaluation. We find that uncertainty metrics generally align with predictive performance, but Bayesian model evidence does so inconsistently. We attribute this to a mismatch in Bayesian and geometric notions of model complexity for the employed last-layer Laplace approximation, and discuss possible remedies. Our findings point towards the potential of uncertainty in guiding symmetry-aware model selection.
comment: 9 pages, 4 figures, 2 tables. In the 8th Workshop on Tractable Probabilistic Modeling at UAI 2025
♻ ☆ On the Effect of Instruction Tuning Loss on Generalization ACL
Instruction Tuning has emerged as a pivotal post-training paradigm that enables pre-trained language models to better follow user instructions. Despite its significance, little attention has been given to optimizing the loss function used. A fundamental, yet often overlooked, question is whether the conventional auto-regressive objective - where loss is computed only on response tokens, excluding prompt tokens - is truly optimal for instruction tuning. In this work, we systematically investigate the impact of differentially weighting prompt and response tokens in instruction tuning loss, and propose Weighted Instruction Tuning (WIT) as a better alternative to conventional instruction tuning. Through extensive experiments on five language models of different families and scale, three finetuning datasets of different sizes, and five diverse evaluation benchmarks, we show that the standard instruction tuning loss often yields suboptimal performance and limited robustness to input prompt variations. We find that a low-to-moderate weight for prompt tokens coupled with a moderate-to-high weight for response tokens yields the best-performing models across settings and also serve as better starting points for the subsequent preference alignment training. These findings highlight the need to reconsider instruction tuning loss and offer actionable insights for developing more robust and generalizable models. Our code is open-sourced at https://github.com/kowndinya-renduchintala/WIT.
comment: To appear in Transactions of the Association for Computational Linguistics (TACL)
♻ ☆ Stylometry recognizes human and LLM-generated texts in short samples
The paper explores stylometry as a method to distinguish between texts created by Large Language Models (LLMs) and humans, addressing issues of model attribution, intellectual property, and ethical AI use. Stylometry has been used extensively to characterise the style and attribute authorship of texts. By applying it to LLM-generated texts, we identify their emergent writing patterns. The paper involves creating a benchmark dataset based on Wikipedia, with (a) human-written term summaries, (b) texts generated purely by LLMs (GPT-3.5/4, LLaMa 2/3, Orca, and Falcon), (c) processed through multiple text summarisation methods (T5, BART, Gensim, and Sumy), and (d) rephrasing methods (Dipper, T5). The 10-sentence long texts were classified by tree-based models (decision trees and LightGBM) using human-designed (StyloMetrix) and n-gram-based (our own pipeline) stylometric features that encode lexical, grammatical, syntactic, and punctuation patterns. The cross-validated results reached a performance of up to .87 Matthews correlation coefficient in the multiclass scenario with 7 classes, and accuracy between .79 and 1. in binary classification, with the particular example of Wikipedia and GPT-4 reaching up to .98 accuracy on a balanced dataset. Shapley Additive Explanations pinpointed features characteristic of the encyclopaedic text type, individual overused words, as well as a greater grammatical standardisation of LLMs with respect to human-written texts. These results show -- crucially, in the context of the increasingly sophisticated LLMs -- that it is possible to distinguish machine- from human-generated texts at least for a well-defined text type.
♻ ☆ Provable Robustness of (Graph) Neural Networks Against Data Poisoning and Backdoor Attacks NeurIPS 2024
Generalization of machine learning models can be severely compromised by data poisoning, where adversarial changes are applied to the training data. This vulnerability has led to interest in certifying (i.e., proving) that such changes up to a certain magnitude do not affect test predictions. We, for the first time, certify Graph Neural Networks (GNNs) against poisoning attacks, including backdoors, targeting the node features of a given graph. Our certificates are white-box and based upon $(i)$ the neural tangent kernel, which characterizes the training dynamics of sufficiently wide networks; and $(ii)$ a novel reformulation of the bilevel optimization problem describing poisoning as a mixed-integer linear program. Consequently, we leverage our framework to provide fundamental insights into the role of graph structure and its connectivity on the worst-case robustness behavior of convolution-based and PageRank-based GNNs. We note that our framework is more general and constitutes the first approach to derive white-box poisoning certificates for NNs, which can be of independent interest beyond graph-related tasks.
comment: Published in TMLR. Best Paper Award at the AdvML-Frontiers @ NeurIPS 2024 workshop. Code available at https://github.com/saper0/qpcert
♻ ☆ A Review of Bayesian Uncertainty Quantification in Deep Probabilistic Image Segmentation
Advances in architectural design, data availability, and compute have driven remarkable progress in semantic segmentation. Yet, these models often rely on relaxed Bayesian assumptions, omitting critical uncertainty information needed for robust decision-making. The resulting reliance on point estimates has fueled interest in probabilistic segmentation, but the literature remains fragmented. In response, this review consolidates and contextualizes foundational concepts in uncertainty modeling, including the non-trivial task of distinguishing between epistemic and aleatoric uncertainty and examining their roles across four key downstream segmentation tasks, highlighting Active Learning as particularly promising. By unifying theory, terminology, and applications, we provide a coherent foundation for researchers and identify critical challenges, such as strong assumptions in spatial aggregation, lack of standardized benchmarks, and pitfalls in current uncertainty quantification methods. We identify trends such as the adoption of contemporary generative models, driven by advances in the broader field of generative modeling, with segmentation-specific innovation primarily in the conditioning mechanisms. Moreover, we observe growing interest in distribution- and sampling-free approaches to uncertainty estimation. We further propose directions for advancing uncertainty-aware segmentation in deep learning, including pragmatic strategies for disentangling different sources of uncertainty, novel uncertainty modeling approaches and improved Transformer-based backbones. In this way, we aim to support the development of more reliable, efficient, and interpretable segmentation models that effectively incorporate uncertainty into real-world applications.
comment: 31 pages of content, revised
♻ ☆ Feature-Based vs. GAN-Based Learning from Demonstrations: When and Why
This survey provides a comparative analysis of feature-based and GAN-based approaches to learning from demonstrations, with a focus on the structure of reward functions and their implications for policy learning. Feature-based methods offer dense, interpretable rewards that excel at high-fidelity motion imitation, yet often require sophisticated representations of references and struggle with generalization in unstructured settings. GAN-based methods, in contrast, use implicit, distributional supervision that enables scalability and adaptation flexibility, but are prone to training instability and coarse reward signals. Recent advancements in both paradigms converge on the importance of structured motion representations, which enable smoother transitions, controllable synthesis, and improved task integration. We argue that the dichotomy between feature-based and GAN-based methods is increasingly nuanced: rather than one paradigm dominating the other, the choice should be guided by task-specific priorities such as fidelity, diversity, interpretability, and adaptability. This work outlines the algorithmic trade-offs and design considerations that underlie method selection, offering a framework for principled decision-making in learning from demonstrations.
♻ ☆ GRAPES: Learning to Sample Graphs for Scalable Graph Neural Networks
Graph neural networks (GNNs) learn to represent nodes by aggregating information from their neighbors. As GNNs increase in depth, their receptive field grows exponentially, leading to high memory costs. Several existing methods address this by sampling a small subset of nodes, scaling GNNs to much larger graphs. These methods are primarily evaluated on homophilous graphs, where neighboring nodes often share the same label. However, most of these methods rely on static heuristics that may not generalize across different graphs or tasks. We argue that the sampling method should be adaptive, adjusting to the complex structural properties of each graph. To this end, we introduce GRAPES, an adaptive sampling method that learns to identify the set of nodes crucial for training a GNN. GRAPES trains a second GNN to predict node sampling probabilities by optimizing the downstream task objective. We evaluate GRAPES on various node classification benchmarks, involving homophilous as well as heterophilous graphs. We demonstrate GRAPES' effectiveness in accuracy and scalability, particularly in multi-label heterophilous graphs. Unlike other sampling methods, GRAPES maintains high accuracy even with smaller sample sizes and, therefore, can scale to massive graphs. Our code is publicly available at https://github.com/dfdazac/grapes.
♻ ☆ EASTER: Embedding Aggregation-based Heterogeneous Models Training in Vertical Federated Learning
Vertical federated learning has garnered significant attention as it allows clients to train machine learning models collaboratively without sharing local data, which protects the client's local private data. However, existing VFL methods face challenges when dealing with heterogeneous local models among participants, which affects optimization convergence and generalization. To address this challenge, this paper proposes a novel approach called Vertical federated learning for training multiple Heterogeneous models (VFedMH). VFedMH focuses on aggregating the local embeddings of each participant's knowledge during forward propagation. To protect the participants' local embedding values, we propose an embedding protection method based on lightweight blinding factors. In particular, participants obtain local embedding using local heterogeneous models. Then the passive party, who owns only features of the sample, injects the blinding factor into the local embedding and sends it to the active party. The active party aggregates local embeddings to obtain global knowledge embeddings and sends them to passive parties. The passive parties then utilize the global embeddings to propagate forward on their local heterogeneous networks. However, the passive party does not own the sample labels, so the local model gradient cannot be calculated locally. To overcome this limitation, the active party assists the passive party in computing its local heterogeneous model gradients. Then, each participant trains their local model using the heterogeneous model gradients. The objective is to minimize the loss value of their respective local heterogeneous models. Extensive experiments are conducted to demonstrate that VFedMH can simultaneously train multiple heterogeneous models with heterogeneous optimization and outperform some recent methods in model performance.
comment: 15 pages, 19 figures
♻ ☆ Fast Fourier Correlation is a Highly Efficient and Accurate Feature Attribution Algorithm from the Perspective of Control Theory and Game Theory
The study of neural networks from the perspective of Fourier features has garnered significant attention. While existing analytical research suggests that neural networks tend to learn low-frequency features, a clear attribution method for identifying the specific learned Fourier features has remained elusive. To bridge this gap, we propose a novel Fourier feature attribution method grounded in signal decomposition theory. Additionally, we analyze the differences between game-theoretic attribution metrics for Fourier and spatial domain features, demonstrating that game-theoretic evaluation metrics are better suited for Fourier-based feature attribution. Our experiments show that Fourier feature attribution exhibits superior feature selection capabilities compared to spatial domain attribution methods. For instance, in the case of Vision Transformers (ViTs) on the ImageNet dataset, only $8\%$ of the Fourier features are required to maintain the original predictions for $80\%$ of the samples. Furthermore, we compare the specificity of features identified by our method against traditional spatial domain attribution methods. Results reveal that Fourier features exhibit greater intra-class concentration and inter-class distinctiveness, indicating their potential for more efficient classification and explainable AI algorithms.
comment: 13 pages, 2 figures
♻ ☆ CLA: Latent Alignment for Online Continual Self-Supervised Learning
Self-supervised learning (SSL) is able to build latent representations that generalize well to unseen data. However, only a few SSL techniques exist for the online CL setting, where data arrives in small minibatches, the model must comply with a fixed computational budget, and task boundaries are absent. We introduce Continual Latent Alignment (CLA), a novel SSL strategy for Online CL that aligns the representations learned by the current model with past representations to mitigate forgetting. We found that our CLA is able to speed up the convergence of the training process in the online scenario, outperforming state-of-the-art approaches under the same computational budget. Surprisingly, we also discovered that using CLA as a pretraining protocol in the early stages of pretraining leads to a better final performance when compared to a full i.i.d. pretraining.
comment: Accepted at CoLLAs 2025 conference (oral)
♻ ☆ TAB: Unified Benchmarking of Time Series Anomaly Detection Methods VLDB2025
Time series anomaly detection (TSAD) plays an important role in many domains such as finance, transportation, and healthcare. With the ongoing instrumentation of reality, more time series data will be available, leading also to growing demands for TSAD. While many TSAD methods already exist, new and better methods are still desirable. However, effective progress hinges on the availability of reliable means of evaluating new methods and comparing them with existing methods. We address deficiencies in current evaluation procedures related to datasets and experimental settings and protocols. Specifically, we propose a new time series anomaly detection benchmark, called TAB. First, TAB encompasses 29 public multivariate datasets and 1,635 univariate time series from different domains to facilitate more comprehensive evaluations on diverse datasets. Second, TAB covers a variety of TSAD methods, including Non-learning, Machine learning, Deep learning, LLM-based, and Time-series pre-trained methods. Third, TAB features a unified and automated evaluation pipeline that enables fair and easy evaluation of TSAD methods. Finally, we employ TAB to evaluate existing TSAD methods and report on the outcomes, thereby offering a deeper insight into the performance of these methods. Besides, all datasets and code are available at https://github.com/decisionintelligence/TAB.
comment: Accepted by PVLDB2025
♻ ☆ PPA-Game: Characterizing and Learning Competitive Dynamics Among Online Content Creators KDD 2026
In this paper, we present the Proportional Payoff Allocation Game (PPA-Game), which characterizes situations where agents compete for divisible resources. In the PPA-game, agents select from available resources, and their payoffs are proportionately determined based on heterogeneous weights attributed to them. Such dynamics simulate content creators on online recommender systems like YouTube and TikTok, who compete for finite consumer attention, with content exposure reliant on inherent and distinct quality. We first conduct a game-theoretical analysis of the PPA-Game. While the PPA-Game does not always guarantee the existence of a pure Nash equilibrium (PNE), we identify prevalent scenarios ensuring its existence. Simulated experiments further prove that the cases where PNE does not exist rarely happen. Beyond analyzing static payoffs, we further discuss the agents' online learning about resource payoffs by integrating a multi-player multi-armed bandit framework. We propose an online algorithm facilitating each agent's maximization of cumulative payoffs over $T$ rounds. Theoretically, we establish that the regret of any agent is bounded by $O(\log^{1 + \eta} T)$ for any $\eta > 0$. Empirical results further validate the effectiveness of our online learning approach.
comment: KDD 2026
♻ ☆ Dynamic Chunking for End-to-End Hierarchical Sequence Modeling
Major progress on language models (LMs) in recent years has largely resulted from moving away from specialized models designed for specific tasks, to general models based on powerful architectures (e.g. the Transformer) that learn everything from raw data. Despite this trend, pre-processing steps such as tokenization remain a barrier to true end-to-end foundation models. We introduce a collection of new techniques that enable a dynamic chunking mechanism which automatically learns content- and context- dependent segmentation strategies learned jointly with the rest of the model. Incorporating this into an explicit hierarchical network (H-Net) allows replacing the (implicitly hierarchical) tokenization-LM-detokenization pipeline with a single model learned fully end-to-end. When compute- and data- matched, an H-Net with one stage of hierarchy operating at the byte level outperforms a strong Transformer language model operating over BPE tokens. Iterating the hierarchy to multiple stages further increases its performance by modeling multiple levels of abstraction, demonstrating significantly better scaling with data and matching the token-based Transformer of twice its size. H-Nets pretrained on English show significantly increased character-level robustness, and qualitatively learn meaningful data-dependent chunking strategies without any heuristics or explicit supervision. Finally, the H-Net's improvement over tokenized pipelines is further increased in languages and modalities with weaker tokenization heuristics, such as Chinese and code, or DNA sequences (nearly 4x improvement in data efficiency over baselines), showing the potential of true end-to-end models that learn and scale better from unprocessed data.
♻ ☆ Context-Aware Deep Lagrangian Networks for Model Predictive Control IROS 2025
Controlling a robot based on physics-consistent dynamic models, such as Deep Lagrangian Networks (DeLaN), can improve the generalizability and interpretability of the resulting behavior. However, in complex environments, the number of objects to potentially interact with is vast, and their physical properties are often uncertain. This complexity makes it infeasible to employ a single global model. Therefore, we need to resort to online system identification of context-aware models that capture only the currently relevant aspects of the environment. While physical principles such as the conservation of energy may not hold across varying contexts, ensuring physical plausibility for any individual context-aware model can still be highly desirable, particularly when using it for receding horizon control methods such as model predictive control (MPC). Hence, in this work, we extend DeLaN to make it context-aware, combine it with a recurrent network for online system identification, and integrate it with an MPC for adaptive, physics-consistent control. We also combine DeLaN with a residual dynamics model to leverage the fact that a nominal model of the robot is typically available. We evaluate our method on a 7-DOF robot arm for trajectory tracking under varying loads. Our method reduces the end-effector tracking error by 39%, compared to a 21% improvement achieved by a baseline that uses an extended Kalman filter.
comment: Accepted to the 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2025)
♻ ☆ Traveling Across Languages: Benchmarking Cross-Lingual Consistency in Multimodal LLMs
The rapid evolution of multimodal large language models (MLLMs) has significantly enhanced their real-world applications. However, achieving consistent performance across languages, especially when integrating cultural knowledge, remains a significant challenge. To better assess this issue, we introduce two new benchmarks: KnowRecall and VisRecall, which evaluate cross-lingual consistency in MLLMs. KnowRecall is a visual question answering benchmark designed to measure factual knowledge consistency in 15 languages, focusing on cultural and historical questions about global landmarks. VisRecall assesses visual memory consistency by asking models to describe landmark appearances in 9 languages without access to images. Experimental results reveal that state-of-the-art MLLMs, including proprietary ones, still struggle to achieve cross-lingual consistency. This underscores the need for more robust approaches that produce truly multilingual and culturally aware models.
comment: https://github.com/nlp-waseda/traveling-across-languages
♻ ☆ LaCoOT: Layer Collapse through Optimal Transport ICCV25
Although deep neural networks are well-known for their outstanding performance in tackling complex tasks, their hunger for computational resources remains a significant hurdle, posing energy-consumption issues and restricting their deployment on resource-constrained devices, preventing their widespread adoption. In this paper, we present an optimal transport-based method to reduce the depth of over-parametrized deep neural networks, alleviating their computational burden. More specifically, we propose a new regularization strategy based on the Max-Sliced Wasserstein distance to minimize the distance between the intermediate feature distributions in the neural network. We show that minimizing this distance enables the complete removal of intermediate layers in the network, achieving better performance/depth trade-off compared to existing techniques. We assess the effectiveness of our method on traditional image classification setups and extend it to generative image models. Our code is available at https://github.com/VGCQ/LaCoOT.
comment: ICCV25
♻ ☆ Tree-Structured Parzen Estimator Can Solve Black-Box Combinatorial Optimization More Efficiently
Tree-structured Parzen estimator (TPE) is a versatile hyperparameter optimization (HPO) method supported by popular HPO tools. Since these HPO tools have been developed in line with the trend of deep learning (DL), the problem setups often used in the DL domain have been discussed for TPE such as multi-objective optimization and multi-fidelity optimization. However, the practical applications of HPO are not limited to DL, and black-box combinatorial optimization is actively utilized in some domains, e.g., chemistry and biology. As combinatorial optimization has been an untouched, yet very important, topic in TPE, we propose an efficient combinatorial optimization algorithm for TPE. In this paper, we first generalize the categorical kernel with the numerical kernel in TPE, enabling us to introduce a distance structure to the categorical kernel. Then we discuss modifications for the newly developed kernel to handle a large combinatorial search space. These modifications reduce the time complexity of the kernel calculation with respect to the size of a combinatorial search space. In the experiments using synthetic problems, we verified that our proposed method identifies better solutions with fewer evaluations than the original TPE. Our algorithm is available in Optuna, an open-source framework for HPO.
comment: Submitted to AutoML Conference
♻ ☆ Comply: Learning Sentences with Complex Weights inspired by Fruit Fly Olfaction
Biologically inspired neural networks offer alternative avenues to model data distributions. FlyVec is a recent example that draws inspiration from the fruit fly's olfactory circuit to tackle the task of learning word embeddings. Surprisingly, this model performs competitively even against deep learning approaches specifically designed to encode text, and it does so with the highest degree of computational efficiency. We pose the question of whether this performance can be improved further. For this, we introduce Comply. By incorporating positional information through complex weights, we enable a single-layer neural network to learn sequence representations. Our experiments show that Comply not only supersedes FlyVec but also performs on par with significantly larger state-of-the-art models. We achieve this without additional parameters. Comply yields sparse contextual representations of sentences that can be interpreted explicitly from the neuron weights.
comment: Accepted at NICE2025
♻ ☆ Generalising Battery Control in Net-Zero Buildings via Personalised Federated RL ICML2025
This work studies the challenge of optimal energy management in building-based microgrids through a collaborative and privacy-preserving framework. We evaluated two common RL algorithms (PPO and TRPO) in different collaborative setups to manage distributed energy resources (DERs) efficiently. Using a customized version of the CityLearn environment and synthetically generated data, we simulate and design net-zero energy scenarios for microgrids composed of multiple buildings. Our approach emphasizes reducing energy costs and carbon emissions while ensuring privacy. Experimental results demonstrate that Federated TRPO is comparable with state-of-the-art federated RL methodologies without hyperparameter tuning. The proposed framework highlights the feasibility of collaborative learning for achieving optimal control policies in energy systems, advancing the goals of sustainable and efficient smart grids. Our code is accessible \href{https://github.com/Optimization-and-Machine-Learning-Lab/energy_fed_trpo.git}{\textit{this repo}}.
comment: Accepted at CO-Build Workshop ICML2025
♻ ☆ Solar Flare Prediction Using Long Short-term Memory (LSTM) and Decomposition-LSTM with Sliding Window Pattern Recognition
We investigate the use of Long Short-Term Memory (LSTM) and Decomposition-LSTM (DLSTM) networks, combined with an ensemble algorithm, to predict solar flare occurrences using time-series data from the GOES catalog. The dataset spans from 2003 to 2023 and includes 151,071 flare events. Among approximately possible patterns, 7,552 yearly pattern windows are identified, highlighting the challenge of long-term forecasting due to the Sun's complex, self-organized criticality-driven behavior. A sliding window technique is employed to detect temporal quasi-patterns in both irregular and regularized flare time series. Regularization reduces complexity, enhances large flare activity, and captures active days more effectively. To address class imbalance, resampling methods are applied. LSTM and DLSTM models are trained on sequences of peak fluxes and waiting times from irregular time series, while LSTM and DLSTM, integrated with an ensemble approach, are applied to sliding windows of regularized time series with a 3-hour interval. Performance metrics, particularly TSS (0.74), recall (0.95) and the area under the curve (AUC=0.87) in the receiver operating characteristic (ROC), indicate that DLSTM with an ensemble approach on regularized time series outperforms other models, offering more accurate large-flare forecasts with fewer false errors compared to models trained on irregular time series. The superior performance of DLSTM is attributed to its ability to decompose time series into trend and seasonal components, effectively isolating random noise. This study underscores the potential of advanced machine learning techniques for solar flare prediction and highlights the importance of incorporating various solar cycle phases and resampling strategies to enhance forecasting reliability.
comment: Published in the Astrophysical Journal Supplement Series, volume 279, 2025, DOI: 10.3847/1538-4365/addc73
♻ ☆ The Price of Freedom: Exploring Expressivity and Runtime Tradeoffs in Equivariant Tensor Products ICML 2025
$E(3)$-equivariant neural networks have demonstrated success across a wide range of 3D modelling tasks. A fundamental operation in these networks is the tensor product, which interacts two geometric features in an equivariant manner to create new features. Due to the high computational complexity of the tensor product, significant effort has been invested to optimize the runtime of this operation. For example, Luo et al. (2024) recently proposed the Gaunt tensor product (GTP) which promises a significant speedup. In this work, we provide a careful, systematic analysis of a number of tensor product operations. In particular, we emphasize that different tensor products are not performing the same operation. The reported speedups typically come at the cost of expressivity. We introduce measures of expressivity and interactability to characterize these differences. In addition, we realized the original implementation of GTP can be greatly simplified by directly using a spherical grid at no cost in asymptotic runtime. This spherical grid approach is faster on our benchmarks and in actual training of the MACE interatomic potential by 30%. Finally, we provide the first systematic microbenchmarks of the various tensor product operations. We find that the theoretical runtime guarantees can differ wildly from empirical performance, demonstrating the need for careful application-specific benchmarking. Code is available at https://github.com/atomicarchitects/PriceofFreedom.
comment: Published at ICML 2025. 27 pages, 10 figures
♻ ☆ (Almost) Free Modality Stitching of Foundation Models
Foundation multi-modal models are often designed by stitching of multiple existing pretrained uni-modal models: for example, an image classifier with an text model. This stitching process is performed by training a connector module that aims to align the representation spaces of these uni-modal models towards a multi-modal objective. However, given the complexity of training such connectors on large scale web-based datasets coupled with the ever-increasing number of available pretrained uni-modal models, the task of uni-modal models selection and subsequent connector module training becomes computationally demanding. To address this under-studied critical problem, we propose Hypernetwork Model Alignment (Hyma), a novel all-in-one solution for optimal uni-modal model selection and connector training by leveraging hypernetworks. Specifically, our framework utilizes the parameter prediction capability of a hypernetwork to obtain jointly trained connector modules for $N \times M$ combinations of uni-modal models. In our experiments, Hyma reduces the cost of searching for the best performing uni-modal model pair by $10\times$, while matching the ranking and trained connector performance obtained via grid search across a suite of diverse multi-modal benchmarks.
comment: Pre-print
♻ ☆ ReVISE: Learning to Refine at Test-Time via Intrinsic Self-Verification ICML 2025
Self-awareness, i.e., the ability to assess and correct one's own generation, is a fundamental aspect of human intelligence, making its replication in large language models (LLMs) an important yet challenging task. Previous works tackle this by employing extensive reinforcement learning or rather relying on large external verifiers. In this work, we propose Refine via Intrinsic Self-Verification (ReVISE), an efficient and effective framework that enables LLMs to self-correct their outputs through self-verification. The core idea of ReVISE is to enable LLMs to verify their reasoning processes and continually rethink reasoning trajectories based on its verification. We introduce a structured curriculum based upon online preference learning to implement this efficiently. Specifically, as ReVISE involves two challenging tasks (i.e., self-verification and reasoning correction), we tackle each task sequentially using curriculum learning, collecting both failed and successful reasoning paths to construct preference pairs for efficient training. During inference, our approach enjoys natural test-time scaling by integrating self-verification and correction capabilities, further enhanced by our proposed confidence-aware decoding mechanism. Our experiments on various reasoning tasks demonstrate that ReVISE achieves efficient self-correction and significantly improves reasoning performance.
comment: Published as conference proceeding for ICML 2025. First two authors contributed equally
♻ ☆ Learning from Label Proportions and Covariate-shifted Instances
In many applications, especially due to lack of supervision or privacy concerns, the training data is grouped into bags of instances (feature-vectors) and for each bag we have only an aggregate label derived from the instance-labels in the bag. In learning from label proportions (LLP) the aggregate label is the average of the instance-labels in a bag, and a significant body of work has focused on training models in the LLP setting to predict instance-labels. In practice however, the training data may have fully supervised albeit covariate-shifted source data, along with the usual target data with bag-labels, and we wish to train a good instance-level predictor on the target domain. We call this the covariate-shifted hybrid LLP problem. Fully supervised covariate shifted data often has useful training signals and the goal is to leverage them for better predictive performance in the hybrid LLP setting. To achieve this, we develop methods for hybrid LLP which naturally incorporate the target bag-labels along with the source instance-labels, in the domain adaptation framework. Apart from proving theoretical guarantees bounding the target generalization error, we also conduct experiments on several publicly available datasets showing that our methods outperform LLP and domain adaptation baselines as well techniques from previous related work.
♻ ☆ Structured Preconditioners in Adaptive Optimization: A Unified Analysis
We present a novel unified analysis for a broad class of adaptive optimization algorithms with structured (e.g., layerwise, diagonal, and kronecker-factored) preconditioners for both online regret minimization and offline convex optimization. Our analysis not only provides matching rate to several important structured preconditioned algorithms including diagonal AdaGrad, full-matrix AdaGrad, and AdaGrad-Norm, but also gives an improved convergence rate for a one-sided variant of Shampoo over that of original Shampoo. Interestingly, more structured preconditioners (e.g., diagonal Adagrad, AdaGrad-Norm which use less space and compute) are often presented as computationally efficient approximations to full-matrix Adagrad, aiming for improved optimization performance through better approximations. Our unified analysis challenges this prevailing view and reveals, perhaps surprisingly, that more structured preconditioners, despite using less space and computation per step, can outperform their less structured counterparts. To demonstrate this, we show that one-sided Shampoo, which is relatively much cheaper than full-matrix AdaGrad could outperform it both theoretically and experimentally.
comment: Fix typos and add remarks for two-sided Shampoo
♻ ☆ MATE: LLM-Powered Multi-Agent Translation Environment for Accessibility Applications
Accessibility remains a critical concern in today's society, as many technologies are not developed to support the full range of user needs. Existing multi-agent systems (MAS) often cannot provide comprehensive assistance for users in need due to the lack of customization stemming from closed-source designs. Consequently, individuals with disabilities frequently encounter significant barriers when attempting to interact with digital environments. We introduce MATE, a multimodal accessibility MAS, which performs the modality conversions based on the user's needs. The system is useful for assisting people with disabilities by ensuring that data will be converted to an understandable format. For instance, if the user cannot see well and receives an image, the system converts this image to its audio description. MATE can be applied to a wide range of domains, industries, and areas, such as healthcare, and can become a useful assistant for various groups of users. The system supports multiple types of models, ranging from LLM API calling to using custom machine learning (ML) classifiers. This flexibility ensures that the system can be adapted to various needs and is compatible with a wide variety of hardware. Since the system is expected to run locally, it ensures the privacy and security of sensitive information. In addition, the framework can be effectively integrated with institutional technologies (e.g., digital healthcare service) for real-time user assistance. Furthermore, we introduce ModCon-Task-Identifier, a model that is capable of extracting the precise modality conversion task from the user input. Numerous experiments show that ModCon-Task-Identifier consistently outperforms other LLMs and statistical models on our custom data. Our code and data are publicly available at https://github.com/AlgazinovAleksandr/Multi-Agent-MATE.
♻ ☆ On the Similarities of Embeddings in Contrastive Learning
Contrastive learning operates on a simple yet effective principle: Embeddings of positive pairs are pulled together, while those of negative pairs are pushed apart. In this paper, we propose a unified framework for understanding contrastive learning through the lens of cosine similarity, and present two key theoretical insights derived from this framework. First, in full-batch settings, we show that perfect alignment of positive pairs is unattainable when negative-pair similarities fall below a threshold, and this misalignment can be mitigated by incorporating within-view negative pairs into the objective. Second, in mini-batch settings, smaller batch sizes induce stronger separation among negative pairs in the embedding space, i.e., higher variance in their similarities, which in turn degrades the quality of learned representations compared to full-batch settings. To address this, we propose an auxiliary loss that reduces the variance of negative-pair similarities in mini-batch settings. Empirical results show that incorporating the proposed loss improves performance in small-batch settings.
comment: contrastive learning, representation learning, embedding, similarity, negative pair, positive pair
♻ ☆ Patch-wise Structural Loss for Time Series Forecasting
Time-series forecasting has gained significant attention in machine learning due to its crucial role in various domains. However, most existing forecasting models rely heavily on point-wise loss functions like Mean Square Error, which treat each time step independently and neglect the structural dependencies inherent in time series data, making it challenging to capture complex temporal patterns accurately. To address these challenges, we propose a novel Patch-wise Structural (PS) loss, designed to enhance structural alignment by comparing time series at the patch level. Through leveraging local statistical properties, such as correlation, variance, and mean, PS loss captures nuanced structural discrepancies overlooked by traditional point-wise losses. Furthermore, it integrates seamlessly with point-wise loss, simultaneously addressing local structural inconsistencies and individual time-step errors. PS loss establishes a novel benchmark for accurately modeling complex time series data and provides a new perspective on time series loss function design. Extensive experiments demonstrate that PS loss significantly improves the performance of state-of-the-art models across diverse real-world datasets.
♻ ☆ Exploring and Improving Initialization for Deep Graph Neural Networks: A Signal Propagation Perspective
Graph Neural Networks (GNNs) often suffer from performance degradation as the network depth increases. This paper addresses this issue by introducing initialization methods that enhance signal propagation (SP) within GNNs. We propose three key metrics for effective SP in GNNs: forward propagation, backward propagation, and graph embedding variation (GEV). While the first two metrics derive from classical SP theory, the third is specifically designed for GNNs. We theoretically demonstrate that a broad range of commonly used initialization methods for GNNs, which exhibit performance degradation with increasing depth, fail to control these three metrics simultaneously. To deal with this limitation, a direct exploitation of the SP analysis--searching for weight initialization variances that optimize the three metrics--is shown to significantly enhance the SP in deep GCNs. This approach is called Signal Propagation on Graph-guided Initialization (SPoGInit). Our experiments demonstrate that SPoGInit outperforms commonly used initialization methods on various tasks and architectures. Notably, SPoGInit enables performance improvements as GNNs deepen, which represents a significant advancement in addressing depth-related challenges and highlights the validity and effectiveness of the SP analysis framework.
comment: Published in TMLR (2025)
♻ ☆ Fully Data-driven but Interpretable Human Behavioural Modelling with Differentiable Discrete Choice Model
Discrete choice models are essential for modelling various decision-making processes in human behaviour. However, the specification of these models has depended heavily on domain knowledge from experts, and the fully automated but interpretable modelling of complex human behaviours has been a long-standing challenge. In this paper, we introduce the differentiable discrete choice model (Diff-DCM), a fully data-driven method for the interpretable modelling, learning, prediction, and control of complex human behaviours, which is realised by differentiable programming. Solely from input features and choice outcomes without any prior knowledge, Diff-DCM can estimate interpretable closed-form utility functions that reproduce observed behaviours. Comprehensive experiments with both synthetic and real-world data demonstrate that Diff-DCM can be applied to various types of data and requires only a small amount of computational resources for the estimations, which can be completed within tens of seconds on a laptop without any accelerators. In these experiments, we also demonstrate that, using its differentiability, Diff-DCM can provide useful insights into human behaviours, such as an optimal intervention path for effective behavioural changes. This study provides a strong basis for the fully automated and reliable modelling, prediction, and control of human behaviours.
♻ ☆ Online Intrinsic Rewards for Decision Making Agents from Large Language Model Feedback
Automatically synthesizing dense rewards from natural language descriptions is a promising paradigm in reinforcement learning (RL), with applications to sparse reward problems, open-ended exploration, and hierarchical skill design. Recent works have made promising steps by exploiting the prior knowledge of large language models (LLMs). However, these approaches suffer from important limitations: they are either not scalable to problems requiring billions of environment samples, due to requiring LLM annotations for each observation, or they require a diverse offline dataset, which may not exist or be impossible to collect. In this work, we address these limitations through a combination of algorithmic and systems-level contributions. We propose ONI, a distributed architecture that simultaneously learns an RL policy and an intrinsic reward function using LLM feedback. Our approach annotates the agent's collected experience via an asynchronous LLM server, which is then distilled into an intrinsic reward model. We explore a range of algorithmic choices for reward modeling with varying complexity, including hashing, classification, and ranking models. Our approach achieves state-of-the-art performance across a range of challenging tasks from the NetHack Learning Environment, while removing the need for large offline datasets required by prior work. We make our code available at https://github.com/facebookresearch/oni .
♻ ☆ BMDetect: A Multimodal Deep Learning Framework for Comprehensive Biomedical Misconduct Detection
Academic misconduct detection in biomedical research remains challenging due to algorithmic narrowness in existing methods and fragmented analytical pipelines. We present BMDetect, a multimodal deep learning framework that integrates journal metadata (SJR, institutional data), semantic embeddings (PubMedBERT), and GPT-4o-mined textual attributes (methodological statistics, data anomalies) for holistic manuscript evaluation. Key innovations include: (1) multimodal fusion of domain-specific features to reduce detection bias; (2) quantitative evaluation of feature importance, identifying journal authority metrics (e.g., SJR-index) and textual anomalies (e.g., statistical outliers) as dominant predictors; and (3) the BioMCD dataset, a large-scale benchmark with 13,160 retracted articles and 53,411 controls. BMDetect achieves 74.33% AUC, outperforming single-modality baselines by 8.6%, and demonstrates transferability across biomedical subfields. This work advances scalable, interpretable tools for safeguarding research integrity.
♻ ☆ Trajectory Imputation in Multi-Agent Sports with Derivative-Accumulating Self-Ensemble ECML
Multi-agent trajectory data collected from domains such as team sports often suffer from missing values due to various factors. While many imputation methods have been proposed for spatiotemporal data, they are not well-suited for multi-agent sports scenarios where player movements are highly dynamic and inter-agent interactions continuously evolve. To address these challenges, we propose MIDAS (Multi-agent Imputer with Derivative-Accumulating Self-ensemble), a framework that imputes multi-agent trajectories with high accuracy and physical plausibility. It jointly predicts positions, velocities, and accelerations through a Set Transformer-based neural network and generates alternative estimates by recursively accumulating predicted velocity and acceleration values. These predictions are then combined using a learnable weighted ensemble to produce final imputed trajectories. Experiments on three sports datasets demonstrate that MIDAS significantly outperforms existing baselines in both positional accuracy and physical plausibility. Lastly, we showcase use cases of MIDAS, such as approximating total distance and pass success probability, to highlight its applicability to practical downstream tasks that require complete tracking data.
comment: Accepted at ECML/PKDD 2025
♻ ☆ Unified ODE Analysis of Smooth Q-Learning Algorithms
Convergence of Q-learning has been the focus of extensive research over the past several decades. Recently, an asymptotic convergence analysis for Q-learning was introduced using a switching system framework. This approach applies the so-called ordinary differential equation (ODE) approach to prove the convergence of the asynchronous Q-learning modeled as a continuous-time switching system, where notions from switching system theory are used to prove its asymptotic stability without using explicit Lyapunov arguments. However, to prove stability, restrictive conditions, such as quasi-monotonicity, must be satisfied for the underlying switching systems, which makes it hard to easily generalize the analysis method to other reinforcement learning algorithms, such as the smooth Q-learning variants. In this paper, we present a more general and unified convergence analysis that improves upon the switching system approach and can analyze Q-learning and its smooth variants. The proposed analysis is motivated by previous work on the convergence of synchronous Q-learning based on $p$-norm serving as a Lyapunov function. However, the proposed analysis addresses more general ODE models that can cover both asynchronous Q-learning and its smooth versions with simpler frameworks.
♻ ☆ Seeding neural network quantum states with tensor network states
We find an efficient approach to approximately convert matrix product states (MPSs) into restricted Boltzmann machine wave functions consisting of a multinomial hidden unit through a canonical polyadic (CP) decomposition of the MPSs. This method allows us to generate well-behaved initial neural network quantum states for quantum many-body ground-state calculations in polynomial time of the number of variational parameters and systematically shorten the distance between the initial states and the ground states with increasing the rank of the CP decomposition. We demonstrate the efficiency of our method by taking the transverse-field Ising model as an example and discuss possible applications of our method to more general quantum many-body systems in which the ground-state wave functions possess complex nodal structures.
comment: 14 pages, 15 figures
♻ ☆ Is Training Data Quality or Quantity More Impactful to Small Language Model Performance?
This study investigates the relative impact of training data quality versus quantity on the performance of small language models (SLMs), utilizing the TinyStories dataset for empirical analysis. Analysis of dataset variations with respect to size (25% and 50% of the original size) and duplication (controlled rates of 25%, 50%, 75%, and 100%) were performed. Model performance was evaluated based on the validation loss, accuracy, and perplexity metrics. Results indicate training data quality plays a more significant role in the overall performance of SLMs, especially given scale of this experiment. Minimal duplication positively impacted model accuracy (+0.87% increase in accuracy at 25% duplication) without significantly increasing perplexity (+0.52% increase going from 0% to 25% duplication) but excessive duplication led to pronounced performance degradation (-40% drop in accuracy at 100% duplication). The implications of this exploration extend beyond just model performance; training large-scale models imposes significant financial and computational burdens, which can be prohibitive for organizations, individuals, and the public at large, especially in developing countries. Additionally, the energy consumption associated with large-scale training raises environmental concerns. Understanding the relative importance of data quality versus quantity could democratize AI technology, making advanced models more accessible and sustainable for all.
comment: 14 pages, 5 tables, 4 figures | Accepted at International Conference on Neural Computing for Advanced Applications 2025, Conference info: https://aaci.org.hk/ncaa2025
Multimedia 4
☆ Teach Me Sign: Stepwise Prompting LLM for Sign Language Production ICIP 2025
Large language models, with their strong reasoning ability and rich knowledge, have brought revolution to many tasks of AI, but their impact on sign language generation remains limited due to its complexity and unique rules. In this paper, we propose TEAch Me Sign (TEAM-Sign), treating sign language as another natural language. By fine-tuning an LLM, we enable it to learn the correspondence between text and sign language, and facilitate generation. Considering the differences between sign and spoken language, we employ a stepwise prompting strategy to extract the inherent sign language knowledge within the LLM, thereby supporting the learning and generation process. Experimental results on How2Sign and Phoenix14T datasets demonstrate that our approach effectively leverages both the sign language knowledge and reasoning capabilities of LLM to align the different distribution and grammatical rules between sign and spoken language.
comment: Accepted by IEEE ICIP 2025
♻ ☆ ICME 2025 Generalizable HDR and SDR Video Quality Measurement Grand Challenge ICME 2025
This paper reports IEEE International Conference on Multimedia \& Expo (ICME) 2025 Grand Challenge on Generalizable HDR and SDR Video Quality Measurement. With the rapid development of video technology, especially High Dynamic Range (HDR) and Standard Dynamic Range (SDR) contents, the need for robust and generalizable Video Quality Assessment (VQA) methods has become increasingly demanded. Existing VQA models often struggle to deliver consistent performance across varying dynamic ranges, distortion types, and diverse content. This challenge was established to benchmark and promote VQA approaches capable of jointly handling HDR and SDR content. In the final evaluation phase, five teams submitted seven models along with technical reports to the Full Reference (FR) and No Reference (NR) tracks. Among them, four methods outperformed VMAF baseline, while the top-performing model achieved state-of-the-art performance, setting a new benchmark for generalizable video quality assessment.
comment: ICME 2025 Grand Challenges
♻ ☆ An End-to-End Pipeline Perspective on Video Streaming in Best-Effort Networks: A Survey and Tutorial
Remaining a dominant force in Internet traffic, video streaming captivates end users, service providers, and researchers. This paper takes a pragmatic approach to reviewing recent advances in the field by focusing on the prevalent streaming paradigm that involves delivering long-form two-dimensional videos over the best-effort Internet with client-side adaptive bitrate (ABR) algorithms and assistance from content delivery networks (CDNs). To enhance accessibility, we supplement the survey with tutorial material. Unlike existing surveys that offer fragmented views, our work provides a holistic perspective on the entire end-to-end streaming pipeline, from video capture by a camera-equipped device to playback by the end user. Our novel perspective covers the ingestion, processing, and distribution stages of the pipeline and addresses key challenges such as video compression, upload, transcoding, ABR algorithms, CDN support, and quality of experience. We review over 200 papers and classify streaming designs by their problem-solving methodology, whether based on intuition (simple heuristics), theory (formal optimization), or machine learning (generalizable data patterns). The survey further refines these methodology-based categories and characterizes each design by additional traits such as compatible codecs and use of super resolution. We connect the reviewed research to real-world applications by discussing the practices of commercial streaming platforms. Finally, the survey highlights prominent current trends and outlines future directions in video streaming.
♻ ☆ QuMATL: Query-based Multi-annotator Tendency Learning
Different annotators often assign different labels to the same sample due to backgrounds or preferences, and such labeling patterns are referred to as tendency. In multi-annotator scenarios, we introduce a novel task called Multi-annotator Tendency Learning (MATL), which aims to capture each annotator tendency. Unlike traditional tasks that prioritize consensus-oriented learning, which averages out annotator differences and leads to tendency information loss, MATL emphasizes learning each annotator tendency, better preserves tendency information. To this end, we propose an efficient baseline method, Query-based Multi-annotator Tendency Learning (QuMATL), which uses lightweight query to represent each annotator for tendency modeling. It saves the costs of building separate conventional models for each annotator, leverages shared learnable queries to capture inter-annotator correlations as an additional hidden supervisory signal to enhance modeling performance. Meanwhile, we provide a new metric, Difference of Inter-annotator Consistency (DIC), to evaluate how effectively models preserve annotators tendency information. Additionally, we contribute two large-scale datasets, STREET and AMER, providing averages of 4300 and 3118 per-annotator labels, respectively. Extensive experiments verified the effectiveness of our QuMATL.
comment: 13 pages
Computation and Language 92
☆ EmbRACE-3K: Embodied Reasoning and Action in Complex Environments
Recent advanced vision-language models(VLMs) have demonstrated strong performance on passive, offline image and video understanding tasks. However, their effectiveness in embodied settings, which require online interaction and active scene understanding remains limited. In such scenarios, an agent perceives the environment from a first-person perspective, with each action dynamically shaping subsequent observations. Even state-of-the-art models such as GPT-4o, Claude 3.5 Sonnet, and Gemini 2.5 Pro struggle in open-environment interactions, exhibiting clear limitations in spatial reasoning and long-horizon planning. To address this gap, we introduce EmRACE-3K, a dataset of over 3,000 language-guided tasks situated in diverse, photorealistic environments constructed using Unreal Engine and the UnrealCV-Zoo framework. The tasks encompass a wide range of embodied challenges, including navigation, object manipulation, and multi-stage goal execution. Each task unfolds as a multi-step trajectory, pairing first-person visual observations with high-level instructions, grounded actions, and natural language rationales that express the agent's intent at every step. Using EmRACE-3K, we establish a benchmark to evaluate the embodied reasoning capabilities of VLMs across three key dimensions: Exploration, Dynamic Spatial-Semantic Reasoning, and Multi-stage Goal Execution. In zero-shot settings, all models achieve success rates below 20%, underscoring the challenge posed by our benchmark and the current limitations of VLMs in interactive environments. To demonstrate the utility of EmRACE-3K, we further fine-tune Qwen2.5-VL-7B using supervised learning followed by reinforcement learning. This approach yields substantial improvements across all three challenge categories, highlighting the dataset's effectiveness in enabling the development of embodied reasoning capabilities.
comment: Project page: https://mxllc.github.io/EmbRACE-3K/
☆ REST: Stress Testing Large Reasoning Models by Asking Multiple Problems at Once
Recent Large Reasoning Models (LRMs) have achieved remarkable progress on task-specific benchmarks, yet their evaluation methods remain constrained by isolated problem-solving paradigms. Existing benchmarks predominantly assess single-question reasoning through sequential testing, resulting critical limitations: (1) vulnerability to data contamination and less challenging (e.g., DeepSeek-R1 achieves 97.0% on MATH500), forcing costly and perpetual creation of new questions with large human efforts, (2) failure to evaluate models under multi-context pressure, a key requirement for real-world deployment. To bridge this gap, we present REST (Reasoning Evaluation through Simultaneous Testing), a stress-testing framework that concurrently exposes LRMs to multiple problems simultaneously. Beyond basic reasoning, REST specifically evaluates several under-tested capabilities: contextual priority allocation, cross-problem interference resistance, and dynamic cognitive load management. Our evaluation reveals several striking findings: Even state-of-the-art (SOTA) models like DeepSeek-R1 exhibit substantial performance degradation under stress testing. Crucially, REST demonstrates stronger discriminative power than existing benchmarks, revealing pronounced performance differences among models that exhibit similar, near-ceiling performance under single-question evaluations. Some key mechanistic insights emerge from our analysis: (1) the "overthinking trap" is a critical factor contributing to the performance degradation; (2) the models trained with "long2short" technique preserve more accuracy of their single-problem performance under REST, outperforming standard-trained counterparts. These results establish REST as a cost-efficient, future-proof evaluation paradigm that better reflects real-world reasoning demands while reducing reliance on continuous human annotation.
comment: REST (Reasoning Evaluation through Simultaneous Testing), a stress-testing framework that concurrently exposes LRMs to multiple problems simultaneously
☆ CodeJudgeBench: Benchmarking LLM-as-a-Judge for Coding Tasks
Large Language Models (LLMs) have significantly advanced the state-of-the-art in various coding tasks. Beyond directly answering user queries, LLMs can also serve as judges, assessing and comparing the quality of responses generated by other models. Such an evaluation capability is crucial both for benchmarking different LLMs and for improving response quality through response ranking. However, despite the growing adoption of the LLM-as-a-Judge paradigm, its effectiveness in coding scenarios remains underexplored due to the absence of dedicated benchmarks. To address this gap, we introduce CodeJudgeBench, a benchmark explicitly designed to evaluate the performance of LLM-as-a-Judge models across three critical coding tasks: code generation, code repair, and unit test generation. Through comprehensive benchmarking of 26 LLM-as-a-Judge models, we find that recent thinking models significantly outperform non-thinking models on our carefully designed code judging tasks. Notably, even relatively small thinking models, such as Qwen3-8B, can outperform specially trained LLM-as-a-Judge models up to 70B in size. Nevertheless, all models still exhibit significant randomness in their judgment of coding tasks. For pairwise judging tasks, simply changing the order in which responses are presented can substantially impact accuracy. In addition, when judging code and unit tests written by different LLMs, LLM-as-a-Judge models also show variance in performance. This sensitivity raises concerns about the reliability and consistency of LLM-as-a-Judge in coding scenarios. Lastly, we study optimal prompting strategies for LLM-as-a-Judge. We find that using pair-wise comparison outperforms scalar point-wise judging. Furthermore, retaining comments and reasoning in the full, unprocessed LLM response leads to improved judge performance.
comment: Dataset is available at https://huggingface.co/datasets/mattymchen/codejudgebench
☆ Reasoning or Memorization? Unreliable Results of Reinforcement Learning Due to Data Contamination
The reasoning capabilities of large language models (LLMs) have been a longstanding focus of research. Recent works have further enhanced these capabilities using reinforcement learning (RL), with many new methods claiming significant improvements with minimal or no external supervision. Surprisingly, some studies even suggest that random or incorrect reward signals can enhance reasoning performance. However, these breakthroughs are mostly reported on the Qwen2.5 model family and evaluated on well-known benchmarks such as MATH-500, AMC, and AIME, while failing to achieve similar gains on other models like Llama, which warrants further investigation. Our analysis shows that although Qwen2.5 achieves strong mathematical reasoning performance, its pretraining on large-scale web corpora makes it vulnerable to data contamination in popular benchmarks. As a result, results derived from these benchmarks may be unreliable. To address this, we introduce a generator that produces fully synthetic arithmetic problems of arbitrary length and difficulty, yielding a clean dataset we call RandomCalculation. Using these leakage-free datasets, we show that only accurate reward signals consistently improve performance, while noisy or incorrect signals do not. We advocate for evaluating RL methods on uncontaminated benchmarks and across diverse model families to ensure trustworthy conclusions.
comment: 26 pages
☆ Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation
Scaling language models unlocks impressive capabilities, but the accompanying computational and memory demands make both training and deployment expensive. Existing efficiency efforts typically target either parameter sharing or adaptive computation, leaving open the question of how to attain both simultaneously. We introduce Mixture-of-Recursions (MoR), a unified framework that combines the two axes of efficiency inside a single Recursive Transformer. MoR reuses a shared stack of layers across recursion steps to achieve parameter efficiency, while lightweight routers enable adaptive token-level thinking by dynamically assigning different recursion depths to individual tokens. This allows MoR to focus quadratic attention computation only among tokens still active at a given recursion depth, further improving memory access efficiency by selectively caching only their key-value pairs. Beyond these core mechanisms, we also propose a KV sharing variant that reuses KV pairs from the first recursion, specifically designed to decrease prefill latency and memory footprint. Across model scales ranging from 135M to 1.7B parameters, MoR forms a new Pareto frontier: at equal training FLOPs and smaller model sizes, it significantly lowers validation perplexity and improves few-shot accuracy, while delivering higher throughput compared with vanilla and existing recursive baselines. These gains demonstrate that MoR is an effective path towards large-model quality without incurring large-model cost.
comment: 36 pages, 9 figures, 14 tables, codes at https://github.com/raymin0223/mixture_of_recursions
☆ DeepResearch$^{\text{Eco}}$: A Recursive Agentic Workflow for Complex Scientific Question Answering in Ecology
We introduce DeepResearch$^{\text{Eco}}$, a novel agentic LLM-based system for automated scientific synthesis that supports recursive, depth- and breadth-controlled exploration of original research questions -- enhancing search diversity and nuance in the retrieval of relevant scientific literature. Unlike conventional retrieval-augmented generation pipelines, DeepResearch enables user-controllable synthesis with transparent reasoning and parameter-driven configurability, facilitating high-throughput integration of domain-specific evidence while maintaining analytical rigor. Applied to 49 ecological research questions, DeepResearch achieves up to a 21-fold increase in source integration and a 14.9-fold rise in sources integrated per 1,000 words. High-parameter settings yield expert-level analytical depth and contextual diversity. Source code available at: https://github.com/sciknoworg/deep-research.
comment: 12 pages, 3 figures
☆ Can You Detect the Difference?
The rapid advancement of large language models (LLMs) has raised concerns about reliably detecting AI-generated text. Stylometric metrics work well on autoregressive (AR) outputs, but their effectiveness on diffusion-based models is unknown. We present the first systematic comparison of diffusion-generated text (LLaDA) and AR-generated text (LLaMA) using 2 000 samples. Perplexity, burstiness, lexical diversity, readability, and BLEU/ROUGE scores show that LLaDA closely mimics human text in perplexity and burstiness, yielding high false-negative rates for AR-oriented detectors. LLaMA shows much lower perplexity but reduced lexical fidelity. Relying on any single metric fails to separate diffusion outputs from human writing. We highlight the need for diffusion-aware detectors and outline directions such as hybrid models, diffusion-specific stylometric signatures, and robust watermarking.
comment: 11 pages, 3 figures, 2 tables. Code and data: https://github.com/ismailtrm/ceng_404. Cross-list requested to cs.AI for AI-safety relevance
☆ MLAR: Multi-layer Large Language Model-based Robotic Process Automation Applicant Tracking
This paper introduces an innovative Applicant Tracking System (ATS) enhanced by a novel Robotic process automation (RPA) framework or as further referred to as MLAR. Traditional recruitment processes often encounter bottlenecks in resume screening and candidate shortlisting due to time and resource constraints. MLAR addresses these challenges employing Large Language Models (LLMs) in three distinct layers: extracting key characteristics from job postings in the first layer, parsing applicant resume to identify education, experience, skills in the second layer, and similarity matching in the third layer. These features are then matched through advanced semantic algorithms to identify the best candidates efficiently. Our approach integrates seamlessly into existing RPA pipelines, automating resume parsing, job matching, and candidate notifications. Extensive performance benchmarking shows that MLAR outperforms the leading RPA platforms, including UiPath and Automation Anywhere, in high-volume resume-processing tasks. When processing 2,400 resumes, MLAR achieved an average processing time of 5.4 seconds per resume, reducing processing time by approximately 16.9% compared to Automation Anywhere and 17.1% compared to UiPath. These results highlight the potential of MLAR to transform recruitment workflows by providing an efficient, accurate, and scalable solution tailored to modern hiring needs.
☆ From BERT to Qwen: Hate Detection across architectures
Online platforms struggle to curb hate speech without over-censoring legitimate discourse. Early bidirectional transformer encoders made big strides, but the arrival of ultra-large autoregressive LLMs promises deeper context-awareness. Whether this extra scale actually improves practical hate-speech detection on real-world text remains unverified. Our study puts this question to the test by benchmarking both model families, classic encoders and next-generation LLMs, on curated corpora of online interactions for hate-speech detection (Hate or No Hate).
comment: 4 pages, 5 figures. EE-559 Deep Learning course project (Group 11)
☆ Referential ambiguity and clarification requests: comparing human and LLM behaviour
In this work we examine LLMs' ability to ask clarification questions in task-oriented dialogues that follow the asynchronous instruction-giver/instruction-follower format. We present a new corpus that combines two existing annotations of the Minecraft Dialogue Corpus -- one for reference and ambiguity in reference, and one for SDRT including clarifications -- into a single common format providing the necessary information to experiment with clarifications and their relation to ambiguity. With this corpus we compare LLM actions with original human-generated clarification questions, examining how both humans and LLMs act in the case of ambiguity. We find that there is only a weak link between ambiguity and humans producing clarification questions in these dialogues, and low correlation between humans and LLMs. Humans hardly ever produce clarification questions for referential ambiguity, but often do so for task-based uncertainty. Conversely, LLMs produce more clarification questions for referential ambiguity, but less so for task uncertainty. We question if LLMs' ability to ask clarification questions is predicated on their recent ability to simulate reasoning, and test this with different reasoning approaches, finding that reasoning does appear to increase question frequency and relevancy.
☆ Multiple Choice Learning of Low Rank Adapters for Language Modeling
We propose LoRA-MCL, a training scheme that extends next-token prediction in language models with a method designed to decode diverse, plausible sentence continuations at inference time. Traditional language modeling is an intrinsically ill-posed problem: given a context, multiple futures may be equally plausible. Our approach leverages Multiple Choice Learning (MCL) and the Winner-Takes-All (WTA) loss to efficiently handle ambiguity through Low-Rank Adaptation (LoRA). We provide a theoretical interpretation of applying Multiple Choice Learning to Language Modeling, assuming the data is generated from a mixture of distributions. To illustrate the proposed approach, we use data sampled from mixtures of Markov chains. We then demonstrate with extensive experiments on real-world visual and audio captioning tasks that our method achieves high diversity and relevance in generated outputs.
☆ Text-to-Remote-Sensing-Image Retrieval beyond RGB Sources
Retrieving relevant imagery from vast satellite archives is crucial for applications like disaster response and long-term climate monitoring. However, most text-to-image retrieval systems are limited to RGB data, failing to exploit the unique physical information captured by other sensors, such as the all-weather structural sensitivity of Synthetic Aperture Radar (SAR) or the spectral signatures in optical multispectral data. To bridge this gap, we introduce CrisisLandMark, a new large-scale corpus of over 647,000 Sentinel-1 SAR and Sentinel-2 multispectral images paired with structured textual annotations for land cover, land use, and crisis events harmonized from authoritative land cover systems (CORINE and Dynamic World) and crisis-specific sources. We then present CLOSP (Contrastive Language Optical SAR Pretraining), a novel framework that uses text as a bridge to align unpaired optical and SAR images into a unified embedding space. Our experiments show that CLOSP achieves a new state-of-the-art, improving retrieval nDGC by 54% over existing models. Additionally, we find that the unified training strategy overcomes the inherent difficulty of interpreting SAR imagery by transferring rich semantic knowledge from the optical domain with indirect interaction. Furthermore, GeoCLOSP, which integrates geographic coordinates into our framework, creates a powerful trade-off between generality and specificity: while the CLOSP excels at general semantic tasks, the GeoCLOSP becomes a specialized expert for retrieving location-dependent crisis events and rare geographic features. This work highlights that the integration of diverse sensor data and geographic context is essential for unlocking the full potential of remote sensing archives.
☆ Devanagari Handwritten Character Recognition using Convolutional Neural Network
Handwritten character recognition is getting popular among researchers because of its possible applications in facilitating technological search engines, social media, recommender systems, etc. The Devanagari script is one of the oldest language scripts in India that does not have proper digitization tools. With the advancement of computing and technology, the task of this research is to extract handwritten Hindi characters from an image of Devanagari script with an automated approach to save time and obsolete data. In this paper, we present a technique to recognize handwritten Devanagari characters using two deep convolutional neural network layers. This work employs a methodology that is useful to enhance the recognition rate and configures a convolutional neural network for effective Devanagari handwritten text recognition (DHTR). This approach uses the Devanagari handwritten character dataset (DHCD), an open dataset with 36 classes of Devanagari characters. Each of these classes has 1700 images for training and testing purposes. This approach obtains promising results in terms of accuracy by achieving 96.36% accuracy in testing and 99.55% in training time.
comment: 9 pages, 6 figures
☆ Meanings are like Onions: a Layered Approach to Metaphor Processing
Metaphorical meaning is not a flat mapping between concepts, but a complex cognitive phenomenon that integrates multiple levels of interpretation. In this paper, we propose a stratified model of metaphor processing that treats meaning as an onion: a multi-layered structure comprising (1) content analysis, (2) conceptual blending, and (3) pragmatic intentionality. This three-dimensional framework allows for a richer and more cognitively grounded approach to metaphor interpretation in computational systems. At the first level, metaphors are annotated through basic conceptual elements. At the second level, we model conceptual combinations, linking components to emergent meanings. Finally, at the third level, we introduce a pragmatic vocabulary to capture speaker intent, communicative function, and contextual effects, aligning metaphor understanding with pragmatic theories. By unifying these layers into a single formal framework, our model lays the groundwork for computational methods capable of representing metaphorical meaning beyond surface associations, toward deeper, more context-sensitive reasoning.
☆ Using AI to replicate human experimental results: a motion study
This paper explores the potential of large language models (LLMs) as reliable analytical tools in linguistic research, focusing on the emergence of affective meanings in temporal expressions involving manner-of-motion verbs. While LLMs like GPT-4 have shown promise across a range of tasks, their ability to replicate nuanced human judgements remains under scrutiny. We conducted four psycholinguistic studies (on emergent meanings, valence shifts, verb choice in emotional contexts, and sentence-emoji associations) first with human participants and then replicated the same tasks using an LLM. Results across all studies show a striking convergence between human and AI responses, with statistical analyses (e.g., Spearman's rho = .73-.96) indicating strong correlations in both rating patterns and categorical choices. While minor divergences were observed in some cases, these did not alter the overall interpretative outcomes. These findings offer compelling evidence that LLMs can augment traditional human-based experimentation, enabling broader-scale studies without compromising interpretative validity. This convergence not only strengthens the empirical foundation of prior human-based findings but also opens possibilities for hypothesis generation and data expansion through AI. Ultimately, our study supports the use of LLMs as credible and informative collaborators in linguistic inquiry.
☆ Bridging Robustness and Generalization Against Word Substitution Attacks in NLP via the Growth Bound Matrix Approach ACL
Despite advancements in Natural Language Processing (NLP), models remain vulnerable to adversarial attacks, such as synonym substitutions. While prior work has focused on improving robustness for feed-forward and convolutional architectures, the robustness of recurrent networks and modern state space models (SSMs), such as S4, remains understudied. These architectures pose unique challenges due to their sequential processing and complex parameter dynamics. In this paper, we introduce a novel regularization technique based on Growth Bound Matrices (GBM) to improve NLP model robustness by reducing the impact of input perturbations on model outputs. We focus on computing the GBM for three architectures: Long Short-Term Memory (LSTM), State Space models (S4), and Convolutional Neural Networks (CNN). Our method aims to (1) enhance resilience against word substitution attacks, (2) improve generalization on clean text, and (3) providing the first systematic analysis of SSM (S4) robustness. Extensive experiments across multiple architectures and benchmark datasets demonstrate that our method improves adversarial robustness by up to 8.8% over existing baselines. These results highlight the effectiveness of our approach, outperforming several state-of-the-art methods in adversarial defense. Codes are available at https://github.com/BouriMohammed/GBM
comment: Accepted to ACL Findings 2025
☆ Grammar-Guided Evolutionary Search for Discrete Prompt Optimisation ECAI 2025
Prompt engineering has proven to be a crucial step in leveraging pretrained large language models (LLMs) in solving various real-world tasks. Numerous solutions have been proposed that seek to automate prompt engineering by using the model itself to edit prompts. However, the majority of state-of-the-art approaches are evaluated on tasks that require minimal prompt templates and on very large and highly capable LLMs. In contrast, solving complex tasks that require detailed information to be included in the prompt increases the amount of text that needs to be optimised. Furthermore, smaller models have been shown to be more sensitive to prompt design. To address these challenges, we propose an evolutionary search approach to automated discrete prompt optimisation consisting of two phases. In the first phase, grammar-guided genetic programming is invoked to synthesise prompt-creating programmes by searching the space of programmes populated by function compositions of syntactic, dictionary-based and LLM-based prompt-editing functions. In the second phase, local search is applied to explore the neighbourhoods of best-performing programmes in an attempt to further fine-tune their performance. Our approach outperforms three state-of-the-art prompt optimisation approaches, PromptWizard, OPRO, and RL-Prompt, on three relatively small general-purpose LLMs in four domain-specific challenging tasks. We also illustrate several examples where these benchmark methods suffer relatively severe performance degradation, while our approach improves performance in almost all task-model combinations, only incurring minimal degradation when it does not.
comment: Accepted for Publication at ECAI 2025
☆ FaceLLM: A Multimodal Large Language Model for Face Understanding ICCV 2025
Multimodal large language models (MLLMs) have shown remarkable performance in vision-language tasks. However, existing MLLMs are primarily trained on generic datasets, limiting their ability to reason on domain-specific visual cues such as those in facial images. In particular, tasks that require detailed understanding of facial structure, expression, emotion, and demographic features remain underexplored by MLLMs due to the lack of large-scale annotated face image-text datasets. In this work, we introduce FaceLLM, a multimodal large language model trained specifically for facial image understanding. To construct the training data, we propose a novel weakly supervised pipeline that uses ChatGPT with attribute-aware prompts to generate high-quality question-answer pairs based on images from the FairFace dataset. The resulting corpus, called FairFaceGPT, covers a diverse set of attributes including expression, pose, skin texture, and forensic information. Our experiments demonstrate that FaceLLM improves the performance of MLLMs on various face-centric tasks and achieves state-of-the-art performance. This work highlights the potential of synthetic supervision via language models for building domain-specialized MLLMs, and sets a precedent for trustworthy, human-centric multimodal AI systems. FairFaceGPT dataset and pretrained FaceLLM models are publicly available in the project page.
comment: Accepted in ICCV 2025 workshops
☆ Absher: A Benchmark for Evaluating Large Language Models Understanding of Saudi Dialects
As large language models (LLMs) become increasingly central to Arabic NLP applications, evaluating their understanding of regional dialects and cultural nuances is essential, particularly in linguistically diverse settings like Saudi Arabia. This paper introduces \texttt{Absher}, a comprehensive benchmark specifically designed to assess LLMs performance across major Saudi dialects. \texttt{Absher} comprises over 18,000 multiple-choice questions spanning six distinct categories: Meaning, True/False, Fill-in-the-Blank, Contextual Usage, Cultural Interpretation, and Location Recognition. These questions are derived from a curated dataset of dialectal words, phrases, and proverbs sourced from various regions of Saudi Arabia. We evaluate several state-of-the-art LLMs, including multilingual and Arabic-specific models. We also provide detailed insights into their capabilities and limitations. Our results reveal notable performance gaps, particularly in tasks requiring cultural inference or contextual understanding. Our findings highlight the urgent need for dialect-aware training and culturally aligned evaluation methodologies to improve LLMs performance in real-world Arabic applications.
☆ Natural Language-based Assessment of L2 Oral Proficiency using LLMs
Natural language-based assessment (NLA) is an approach to second language assessment that uses instructions - expressed in the form of can-do descriptors - originally intended for human examiners, aiming to determine whether large language models (LLMs) can interpret and apply them in ways comparable to human assessment. In this work, we explore the use of such descriptors with an open-source LLM, Qwen 2.5 72B, to assess responses from the publicly available S&I Corpus in a zero-shot setting. Our results show that this approach - relying solely on textual information - achieves competitive performance: while it does not outperform state-of-the-art speech LLMs fine-tuned for the task, it surpasses a BERT-based model trained specifically for this purpose. NLA proves particularly effective in mismatched task settings, is generalisable to other data types and languages, and offers greater interpretability, as it is grounded in clearly explainable, widely applicable language descriptors.
comment: Accepted for the 10th Workshop on Speech and Language Technology in Education (SLaTE 2025)
☆ Abusive text transformation using LLMs
Although Large Language Models (LLMs) have demonstrated significant advancements in natural language processing tasks, their effectiveness in the classification and transformation of abusive text into non-abusive versions remains an area for exploration. In this study, we aim to use LLMs to transform abusive text (tweets and reviews) featuring hate speech and swear words into non-abusive text, while retaining the intent of the text. We evaluate the performance of two state-of-the-art LLMs, such as Gemini, GPT-4o, DeekSeek and Groq, on their ability to identify abusive text. We them to transform and obtain a text that is clean from abusive and inappropriate content but maintains a similar level of sentiment and semantics, i.e. the transformed text needs to maintain its message. Afterwards, we evaluate the raw and transformed datasets with sentiment analysis and semantic analysis. Our results show Groq provides vastly different results when compared with other LLMs. We have identified similarities between GPT-4o and DeepSeek-V3.
☆ Task-Based Flexible Feature Distillation for LLMs
Knowledge Distillation (KD) in general and feature distillation in particular are promising techniques for reducing the high computational demand of large language models (LLMs). However, traditional feature KD methods typically assume that the teacher and the student share the same hidden size, limiting the flexibility of the student's architecture. A common solution to this problem involves training a linear projector to align their feature spaces, but this introduces additional parameters that must be learned from scratch and often degrades performance on downstream tasks, especially in generative settings. To address this issue, in this work, we propose a novel task-based feature distillation method that enables knowledge transfer between teacher and student models with different hidden layer dimensions, without introducing any new parameters. Leveraging the insight that only a subset of LLM components contribute significantly to a specific downstream task, our approach identifies the most task-relevant hidden units in the teacher and directly distills their activations to the student. Our method is flexible and easily integrates with other distillation frameworks. Empirical results show consistent improvements over prior approaches across diverse tasks, including classification, instruction-following, and summarization, achieving up to a 3\% performance gain over the linear projection baseline.
☆ Fusing Large Language Models with Temporal Transformers for Time Series Forecasting
Recently, large language models (LLMs) have demonstrated powerful capabilities in performing various tasks and thus are applied by recent studies to time series forecasting (TSF) tasks, which predict future values with the given historical time series. Existing LLM-based approaches transfer knowledge learned from text data to time series prediction using prompting or fine-tuning strategies. However, LLMs are proficient at reasoning over discrete tokens and semantic patterns but are not initially designed to model continuous numerical time series data. The gaps between text and time series data lead LLMs to achieve inferior performance to a vanilla Transformer model that is directly trained on TSF data. However, the vanilla Transformers often struggle to learn high-level semantic patterns. In this paper, we design a novel Transformer-based architecture that complementarily leverages LLMs and vanilla Transformers, so as to integrate the high-level semantic representations learned by LLMs into the temporal information encoded by time series Transformers, where a hybrid representation is obtained by fusing the representations from the LLM and the Transformer. The resulting fused representation contains both historical temporal dynamics and semantic variation patterns, allowing our model to predict more accurate future values. Experiments on benchmark datasets demonstrate the effectiveness of the proposed approach.
☆ Enhancing Chain-of-Thought Reasoning with Critical Representation Fine-tuning ACL 2025
Representation Fine-tuning (ReFT), a recently proposed Parameter-Efficient Fine-Tuning (PEFT) method, has attracted widespread attention for significantly improving parameter efficiency by editing representation space alone. In this work, we investigate applying ReFT to complex reasoning tasks. However, directly using the native ReFT method, which modifies fixed representations at the beginning and end of each layer, yields suboptimal performance, as these fixed-position representations have uncertain impact on the outputs. We observe that, in complex reasoning tasks, there often exist certain critical representations. These representations either integrate significant information from preceding layers or regulate subsequent layer representations. Through layer-by-layer propagation, they exert a substantial influence on the final output. Naturally, fine-tuning these critical representations has the potential to greatly enhance reasoning performance. Building upon these insights, we propose Critical Representation Fine-Tuning (CRFT), a novel method that identifies and optimizes these critical representations through information flow analysis. CRFT operates within a supervised learning framework, dynamically optimizing critical representations in a low-rank linear subspace while freezing the base model. The effectiveness and efficiency of our method are validated across eight benchmarks for arithmetic and commonsense reasoning, using LLaMA and Mistral model families. Furthermore, our method also adapts effectively to few-shot settings, boosting one-shot accuracy by 16.4%. Our work highlights the untapped potential of representation-level optimization for CoT reasoning, offering a lightweight yet powerful alternative to traditional PEFT methods.
comment: Accepted by ACL 2025
☆ Cultural Bias in Large Language Models: Evaluating AI Agents through Moral Questionnaires
Are AI systems truly representing human values, or merely averaging across them? Our study suggests a concerning reality: Large Language Models (LLMs) fail to represent diverse cultural moral frameworks despite their linguistic capabilities. We expose significant gaps between AI-generated and human moral intuitions by applying the Moral Foundations Questionnaire across 19 cultural contexts. Comparing multiple state-of-the-art LLMs' origins against human baseline data, we find these models systematically homogenize moral diversity. Surprisingly, increased model size doesn't consistently improve cultural representation fidelity. Our findings challenge the growing use of LLMs as synthetic populations in social science research and highlight a fundamental limitation in current AI alignment approaches. Without data-driven alignment beyond prompting, these systems cannot capture the nuanced, culturally-specific moral intuitions. Our results call for more grounded alignment objectives and evaluation metrics to ensure AI systems represent diverse human values rather than flattening the moral landscape.
comment: 15pages, 1 figure, 2 tables
☆ GeLaCo: An Evolutionary Approach to Layer Compression
Large Language Models (LLM) have achieved remarkable performance across a large number of tasks, but face critical deployment and usage barriers due to substantial computational requirements. Model compression methods, which aim to reduce model size while preserving its capacity, are an important means to mitigate these issues. Promising approaches along these lines, such as structured pruning, typically require costly empirical search for optimal variants and may run the risk of ignoring better solutions. In this work we introduce GeLaCo, an evolutionary approach to LLM compression via layer collapse. Our approach supports an efficient exploration of the compression solution space via population-based search and a module-wise similarity fitness function capturing attention, feed-forward, and hidden state representations. GeLaCo also supports both single and multi-objective evolutionary compression search, establishing the first Pareto frontier along compression and quality axes. We evaluate GeLaCo solutions via both perplexity-based and generative evaluations over foundational and instruction-tuned models, outperforming state-of-the-art alternatives.
☆ PRISM: Fine-Grained Paper-to-Paper Retrieval with Multi-Aspect-Aware Query Optimization
Scientific paper retrieval, particularly framed as document-to-document retrieval, aims to identify relevant papers in response to a long-form query paper, rather than a short query string. Previous approaches to this task have focused on abstracts, embedding them into dense vectors as surrogates for full documents and calculating similarity across them, although abstracts provide only sparse and high-level summaries. To address this, we propose PRISM, a novel document-to-document retrieval method that introduces multiple, fine-grained representations for both the query and candidate papers. In particular, each query paper is decomposed into multiple aspect-specific views and individually embedded, which are then matched against candidate papers similarity segmented to consider their multifaceted dimensions. Moreover, we present SciFullBench, a novel benchmark in which the complete and segmented context of full papers for both queries and candidates is available. Then, experimental results show that PRISM improves performance by an average of 4.3% over existing retrieval baselines.
☆ Automating SPARQL Query Translations between DBpedia and Wikidata
This paper investigates whether state-of-the-art Large Language Models (LLMs) can automatically translate SPARQL between popular Knowledge Graph (KG) schemas. We focus on translations between the DBpedia and Wikidata KG, and later on DBLP and OpenAlex KG. This study addresses a notable gap in KG interoperability research by rigorously evaluating LLM performance on SPARQL-to-SPARQL translation. Two benchmarks are assembled, where the first align 100 DBpedia-Wikidata queries from QALD-9-Plus; the second contains 100 DBLP queries aligned to OpenAlex, testing generalizability beyond encyclopaedic KGs. Three open LLMs: Llama-3-8B, DeepSeek-R1-Distill-Llama-70B, and Mistral-Large-Instruct-2407 are selected based on their sizes and architectures and tested with zero-shot, few-shot, and two chain-of-thought variants. Outputs were compared with gold answers, and resulting errors were categorized. We find that the performance varies markedly across models and prompting strategies, and that translations for Wikidata to DBpedia work far better than translations for DBpedia to Wikidata.
comment: 18 pages, 2 figues. Paper accepted at SEMANTiCS 2025 conference happening on September 2025
☆ Cross-modal Associations in Vision and Language Models: Revisiting the bouba-kiki effect
Recent advances in multimodal models have raised questions about whether vision-and-language models (VLMs) integrate cross-modal information in ways that reflect human cognition. One well-studied test case in this domain is the bouba-kiki effect, where humans reliably associate pseudowords like "bouba" with round shapes and "kiki" with jagged ones. Given the mixed evidence found in prior studies for this effect in VLMs, we present a comprehensive re-evaluation focused on two variants of CLIP, ResNet and Vision Transformer (ViT), given their centrality in many state-of-the-art VLMs. We apply two complementary methods closely modelled after human experiments: a prompt-based evaluation that uses probabilities as model preference, and we use Grad-CAM as a novel way to interpret visual attention in shape-word matching tasks. Our findings show that these models do not consistently exhibit the bouba-kiki effect. While ResNet shows a preference for round shapes, overall performance across both models lacks the expected associations. Moreover, direct comparison with prior human data on the same task shows that the models' responses fall markedly short of the robust, modality-integrated behaviour characteristic of human cognition. These results contribute to the ongoing debate about the extent to which VLMs truly understand cross-modal concepts, highlighting limitations in their internal representations and alignment with human intuitions.
☆ Protective Factor-Aware Dynamic Influence Learning for Suicide Risk Prediction on Social Media
Suicide is a critical global health issue that requires urgent attention. Even though prior work has revealed valuable insights into detecting current suicide risk on social media, little attention has been paid to developing models that can predict subsequent suicide risk over time, limiting their ability to capture rapid fluctuations in individuals' mental state transitions. In addition, existing work ignores protective factors that play a crucial role in suicide risk prediction, focusing predominantly on risk factors alone. Protective factors such as social support and coping strategies can mitigate suicide risk by moderating the impact of risk factors. Therefore, this study proposes a novel framework for predicting subsequent suicide risk by jointly learning the dynamic influence of both risk factors and protective factors on users' suicide risk transitions. We propose a novel Protective Factor-Aware Dataset, which is built from 12 years of Reddit posts along with comprehensive annotations of suicide risk and both risk and protective factors. We also introduce a Dynamic Factors Influence Learning approach that captures the varying impact of risk and protective factors on suicide risk transitions, recognizing that suicide risk fluctuates over time according to established psychological theories. Our thorough experiments demonstrate that the proposed model significantly outperforms state-of-the-art models and large language models across three datasets. In addition, the proposed Dynamic Factors Influence Learning provides interpretable weights, helping clinicians better understand suicidal patterns and enabling more targeted intervention strategies.
☆ On The Role of Intentionality in Knowledge Representation: Analyzing Scene Context for Cognitive Agents with a Tiny Language Model
Since Searle's work deconstructing intent and intentionality in the realm of philosophy, the practical meaning of intent has received little attention in science and technology. Intentionality and context are both central to the scope of Promise Theory's model of Semantic Spacetime, used as an effective Tiny Language Model. One can identify themes and concepts from a text, on a low level (without knowledge of the specific language) by using process coherence as a guide. Any agent process can assess superficially a degree of latent `intentionality' in data by looking for anomalous multi-scale anomalies and assessing the work done to form them. Scale separation can be used to sort parts into `intended' content and `ambient context', using the spacetime coherence as a measure. This offers an elementary but pragmatic interpretation of latent intentionality for very low computational cost, and without reference to extensive training or reasoning capabilities. The process is well within the reach of basic organisms as it does not require large scale artificial probabilistic batch processing. The level of concept formation depends, however, on the memory capacity of the agent.
☆ TextOmics-Guided Diffusion for Hit-like Molecular Generation
Hit-like molecular generation with therapeutic potential is essential for target-specific drug discovery. However, the field lacks heterogeneous data and unified frameworks for integrating diverse molecular representations. To bridge this gap, we introduce TextOmics, a pioneering benchmark that establishes one-to-one correspondences between omics expressions and molecular textual descriptions. TextOmics provides a heterogeneous dataset that facilitates molecular generation through representations alignment. Built upon this foundation, we propose ToDi, a generative framework that jointly conditions on omics expressions and molecular textual descriptions to produce biologically relevant, chemically valid, hit-like molecules. ToDi leverages two encoders (OmicsEn and TextEn) to capture multi-level biological and semantic associations, and develops conditional diffusion (DiffGen) for controllable generation. Extensive experiments confirm the effectiveness of TextOmics and demonstrate ToDi outperforms existing state-of-the-art approaches, while also showcasing remarkable potential in zero-shot therapeutic molecular generation. Sources are available at: https://github.com/hala-ToDi.
☆ Tiny Reward Models ICML
Large decoder-based language models have become the dominant architecture for reward modeling in reinforcement learning from human feedback (RLHF). However, as reward models are increasingly deployed in test-time strategies, their inference costs become a growing concern. We present TinyRM, a family of small, bidirectional masked language models (MLMs) with as few as 400 million parameters, that rival the capabilities of models over 175 times larger on reasoning and safety preference modeling tasks. TinyRM combines FLAN-style prompting, Directional Low-Rank Adaptation (DoRA), and layer freezing to achieve strong performance on RewardBench, despite using significantly fewer resources. Our experiments suggest that small models benefit from domain-specific tuning strategies, particularly in reasoning, where lightweight finetuning methods are especially effective. While challenges remain in building generalist models and conversational preference modeling, our preliminary results highlight the promise of lightweight bidirectional architectures as efficient, scalable alternatives for preference modeling.
comment: 2025 ICML Efficient Systems for Foundation Models Workshop
☆ Enhancing Retrieval Augmented Generation with Hierarchical Text Segmentation Chunking
Retrieval-Augmented Generation (RAG) systems commonly use chunking strategies for retrieval, which enhance large language models (LLMs) by enabling them to access external knowledge, ensuring that the retrieved information is up-to-date and domain-specific. However, traditional methods often fail to create chunks that capture sufficient semantic meaning, as they do not account for the underlying textual structure. This paper proposes a novel framework that enhances RAG by integrating hierarchical text segmentation and clustering to generate more meaningful and semantically coherent chunks. During inference, the framework retrieves information by leveraging both segment-level and cluster-level vector representations, thereby increasing the likelihood of retrieving more precise and contextually relevant information. Evaluations on the NarrativeQA, QuALITY, and QASPER datasets indicate that the proposed method achieved improved results compared to traditional chunking techniques.
☆ MixLoRA-DSI: Dynamically Expandable Mixture-of-LoRA Experts for Rehearsal-Free Generative Retrieval over Dynamic Corpora
Continually updating model-based indexes in generative retrieval with new documents remains challenging, as full retraining is computationally expensive and impractical under resource constraints. We propose MixLoRA-DSI, a novel framework that combines an expandable mixture of Low-Rank Adaptation experts with a layer-wise out-of-distribution (OOD)-driven expansion strategy. Instead of allocating new experts for each new corpus, our proposed expansion strategy enables sublinear parameter growth by selectively introducing new experts only when significant number of OOD documents are detected. Experiments on NQ320k and MS MARCO Passage demonstrate that MixLoRA-DSI outperforms full-model update baselines, with minimal parameter overhead and substantially lower training costs.
☆ ViTCoT: Video-Text Interleaved Chain-of-Thought for Boosting Video Understanding in Large Language Models ACM MM 2025
Video understanding plays a vital role in bridging low-level visual signals with high-level cognitive reasoning, and is fundamental to applications such as autonomous driving, embodied AI, and the broader pursuit of AGI. The rapid development of large language models (LLMs), particularly those utilizing Chain-of-Thought (CoT) technology, has significantly advanced video reasoning capabilities. However, current approaches primarily depend on textual information for reasoning, overlooking the visual modality in the actual video reasoning process. In contrast, humans naturally re-examine visual content while reasoning. Motivated by this, we introduce a novel video reasoning paradigm: Video-Text Interleaved CoT (ViTCoT), which facilitates more intuitive and cognitively aligned reasoning. To the end, first, we construct the Video-Text Interleaved Benchmark (ViTIB), which is created using MLLMs for key-video selection and manually verified. Furthermore, we extensively explore the potential of the ViTCoT paradigm in the video understanding field. Extensive experiments demonstrate that ViTCoT significantly enhances performance compared to the traditional text-only CoT paradigm and effectively activates more neuron values in MLLMs.
comment: Accepted by ACM MM 2025
☆ Function Induction and Task Generalization: An Interpretability Study with Off-by-One Addition SC
Large language models demonstrate the intriguing ability to perform unseen tasks via in-context learning. However, it remains unclear what mechanisms inside the model drive such task-level generalization. In this work, we approach this question through the lens of off-by-one addition (i.e., 1+1=3, 2+2=5, 3+3=?), a two-step, counterfactual task with an unexpected +1 function as a second step. Leveraging circuit-style interpretability techniques such as path patching, we analyze the models' internal computations behind their notable performance and present three key findings. First, we uncover a function induction mechanism that explains the model's generalization from standard addition to off-by-one addition. This mechanism resembles the structure of the induction head mechanism found in prior work and elevates it to a higher level of abstraction. Second, we show that the induction of the +1 function is governed by multiple attention heads in parallel, each of which emits a distinct piece of the +1 function. Finally, we find that this function induction mechanism is reused in a broader range of tasks, including synthetic tasks such as shifted multiple-choice QA and algorithmic tasks such as base-8 addition. Overall, our findings offer deeper insights into how reusable and composable structures within language models enable task-level generalization.
comment: Code: https://github.com/INK-USC/function-induction
☆ MultiVox: Benchmarking Voice Assistants for Multimodal Interactions
The rapid progress of Large Language Models (LLMs) has empowered omni models to act as voice assistants capable of understanding spoken dialogues. These models can process multimodal inputs beyond text, such as speech and visual data, enabling more context-aware interactions. However, current benchmarks fall short in comprehensively evaluating how well these models generate context-aware responses, particularly when it comes to implicitly understanding fine-grained speech characteristics, such as pitch, emotion, timbre, and volume or the environmental acoustic context such as background sounds. Additionally, they inadequately assess the ability of models to align paralinguistic cues with complementary visual signals to inform their responses. To address these gaps, we introduce MultiVox, the first omni voice assistant benchmark designed to evaluate the ability of voice assistants to integrate spoken and visual cues including paralinguistic speech features for truly multimodal understanding. Specifically, MultiVox includes 1000 human-annotated and recorded speech dialogues that encompass diverse paralinguistic features and a range of visual cues such as images and videos. Our evaluation on 9 state-of-the-art models reveals that, although humans excel at these tasks, current models consistently struggle to produce contextually grounded responses.
comment: Work In Progress
☆ LLMs on Trial: Evaluating Judicial Fairness for Large Language Models
Large Language Models (LLMs) are increasingly used in high-stakes fields where their decisions impact rights and equity. However, LLMs' judicial fairness and implications for social justice remain underexplored. When LLMs act as judges, the ability to fairly resolve judicial issues is a prerequisite to ensure their trustworthiness. Based on theories of judicial fairness, we construct a comprehensive framework to measure LLM fairness, leading to a selection of 65 labels and 161 corresponding values. Applying this framework to the judicial system, we compile an extensive dataset, JudiFair, comprising 177,100 unique case facts. To achieve robust statistical inference, we develop three evaluation metrics, inconsistency, bias, and imbalanced inaccuracy, and introduce a method to assess the overall fairness of multiple LLMs across various labels. Through experiments with 16 LLMs, we uncover pervasive inconsistency, bias, and imbalanced inaccuracy across models, underscoring severe LLM judicial unfairness. Particularly, LLMs display notably more pronounced biases on demographic labels, with slightly less bias on substance labels compared to procedure ones. Interestingly, increased inconsistency correlates with reduced biases, but more accurate predictions exacerbate biases. While we find that adjusting the temperature parameter can influence LLM fairness, model size, release date, and country of origin do not exhibit significant effects on judicial fairness. Accordingly, we introduce a publicly available toolkit containing all datasets and code, designed to support future research in evaluating and improving LLM fairness.
☆ Testing Hypotheses from the Social Approval Theory of Online Hate: An Analysis of 110 Million Posts from Parler
In this paper, we explored how online hate is motivated by receiving social approval from others. We specifically examined two central tenets of Walther's (2024) social approval theory of online hate: (H1a) more signals of social approval on hate messages predicts more subsequent hate messages, and (H1b) as social approval increases, hate speech messages become more extreme. Using over 110 million posts from Parler (2018-2021), we observed that the number of upvotes a person received on a hate speech post was unassociated with the amount of hate speech in their next post and posts during the next week, month, three months, and six months. Between-person effects revealed an average negative relationship between social approval and hate speech production at the post level, but this relationship was mixed at other time intervals. Social approval reinforcement mechanisms of online hate may operate differently on niche social media platforms.
☆ Automated Thematic Analyses Using LLMs: Xylazine Wound Management Social Media Chatter Use Case
Background Large language models (LLMs) face challenges in inductive thematic analysis, a task requiring deep interpretive and domain-specific expertise. We evaluated the feasibility of using LLMs to replicate expert-driven thematic analysis of social media data. Methods Using two temporally non-intersecting Reddit datasets on xylazine (n=286 and n=686, for model optimization and validation, respectively) with twelve expert-derived themes, we evaluated five LLMs against expert coding. We modeled the task as a series of binary classifications, rather than a single, multi-label classification, employing zero-, single-, and few-shot prompting strategies and measuring performance via accuracy, precision, recall, and F1-score. Results On the validation set, GPT-4o with two-shot prompting performed best (accuracy: 90.9%; F1-score: 0.71). For high-prevalence themes, model-derived thematic distributions closely mirrored expert classifications (e.g., xylazine use: 13.6% vs. 17.8%; MOUD use: 16.5% vs. 17.8%). Conclusions Our findings suggest that few-shot LLM-based approaches can automate thematic analyses, offering a scalable supplement for qualitative research. Keywords: thematic analysis, large language models, natural language processing, qualitative analysis, social media, prompt engineering, public health
comment: Pages: 19, Abstract word count: 151 words, Manuscript word count: 2185 words, References: 14, Figures: 3, Tables: 2
☆ Can Multimodal Foundation Models Understand Schematic Diagrams? An Empirical Study on Information-Seeking QA over Scientific Papers ACL 2025
This paper introduces MISS-QA, the first benchmark specifically designed to evaluate the ability of models to interpret schematic diagrams within scientific literature. MISS-QA comprises 1,500 expert-annotated examples over 465 scientific papers. In this benchmark, models are tasked with interpreting schematic diagrams that illustrate research overviews and answering corresponding information-seeking questions based on the broader context of the paper. We assess the performance of 18 frontier multimodal foundation models, including o4-mini, Gemini-2.5-Flash, and Qwen2.5-VL. We reveal a significant performance gap between these models and human experts on MISS-QA. Our analysis of model performance on unanswerable questions and our detailed error analysis further highlight the strengths and limitations of current models, offering key insights to enhance models in comprehending multimodal scientific literature.
comment: ACL 2025 Findings
☆ Theory of Mind and Self-Disclosure to CUIs
Self-disclosure is important to help us feel better, yet is often difficult. This difficulty can arise from how we think people are going to react to our self-disclosure. In this workshop paper, we briefly discuss self-disclosure to conversational user interfaces (CUIs) in relation to various social cues. We then, discuss how expressions of uncertainty or representation of a CUI's reasoning could help encourage self-disclosure, by making a CUI's intended "theory of mind" more transparent to users.
comment: Workshop paper presented at ToMinHAI at CUI'2025: Theory of Mind in Human-CUI Interaction, held in conjunction with the 2025 ACM conference on Conversational User Interfaces, July 8th, 2025. 4 pages. 3 figures
☆ Applying Text Embedding Models for Efficient Analysis in Labeled Property Graphs
Labeled property graphs often contain rich textual attributes that can enhance analytical tasks when properly leveraged. This work explores the use of pretrained text embedding models to enable efficient semantic analysis in such graphs. By embedding textual node and edge properties, we support downstream tasks including node classification and relation prediction with improved contextual understanding. Our approach integrates language model embeddings into the graph pipeline without altering its structure, demonstrating that textual semantics can significantly enhance the accuracy and interpretability of property graph analysis.
☆ Language Models for Adult Service Website Text Analysis
Sex trafficking refers to the use of force, fraud, or coercion to compel an individual to perform in commercial sex acts against their will. Adult service websites (ASWs) have and continue to be linked to sex trafficking, offering a platform for traffickers to advertise their victims. Thus, organizations involved in the fight against sex trafficking often use ASW data when attempting to identify potential sex trafficking victims. A critical challenge in transforming ASW data into actionable insight is text analysis. Previous research using ASW data has shown that ASW ad text is important for linking ads. However, working with this text is challenging due to its extensive use of emojis, poor grammar, and deliberate obfuscation to evade law enforcement scrutiny. We conduct a comprehensive study of language modeling approaches for this application area, including simple information retrieval methods, pre-trained transformers, and custom transformer models. We demonstrate that characteristics of ASW text data allow efficient custom transformer models to be trained with relatively small GPU resources and used efficiently for inference on consumer hardware. Our custom models outperform fine-tuned variants of well-known encoder-only transformer models, including BERT-base, RoBERTa, and ModernBERT, on accuracy, recall, F1 score, and ROC AUC. We demonstrate the use of our best-performing custom configuration on three tasks related to ASW data analysis: (i) decomposing the giant component in a graph representation of ASW data, (ii) clustering ASW ad text, and (iii) using the learned token embeddings to understand the use of emojis in the illicit context we study. The models we develop represent a significant advancement in ASW text analysis, which can be leveraged in a variety of downstream applications and research.
comment: 32 pages, 12 figures, 1 table
☆ From Semantic Web and MAS to Agentic AI: A Unified Narrative of the Web of Agents
The concept of the Web of Agents (WoA), which transforms the static, document-centric Web into an environment of autonomous agents acting on users' behalf, has attracted growing interest as large language models (LLMs) become more capable. However, research in this area is still fragmented across different communities. Contemporary surveys catalog the latest LLM-powered frameworks, while the rich histories of Multi-Agent Systems (MAS) and the Semantic Web are often treated as separate, legacy domains. This fragmentation obscures the intellectual lineage of modern systems and hinders a holistic understanding of the field's trajectory. We present the first comprehensive evolutionary overview of the WoA. We show that modern protocols like A2A and the MCP, are direct evolutionary responses to the well-documented limitations of earlier standards like FIPA standards and OWL-based semantic agents. To systematize this analysis, we introduce a four-axis taxonomy (semantic foundation, communication paradigm, locus of intelligence, discovery mechanism). This framework provides a unified analytical lens for comparing agent architectures across all generations, revealing a clear line of descent where others have seen a disconnect. Our analysis identifies a paradigm shift in the 'locus of intelligence': from being encoded in external data (Semantic Web) or the platform (MAS) to being embedded within the agent's core model (LLM). This shift is foundational to modern Agentic AI, enabling the scalable and adaptive systems the WoA has long envisioned. We conclude that while new protocols are essential, they are insufficient for building a robust, open, trustworthy ecosystem. Finally, we argue that the next research frontier lies in solving persistent socio-technical challenges, and we map out a new agenda focused on decentralized identity, economic models, security, and governance for the emerging WoA.
comment: 33 pages, 9 figures, 8 tables
☆ HKGAI-V1: Towards Regional Sovereign Large Language Model for Hong Kong
This paper presents the development of HKGAI-V1, a foundational sovereign large language model (LLM), developed as part of an initiative to establish value-aligned AI infrastructure specifically tailored for Hong Kong. Addressing the region's unique multilingual environment (Cantonese, Mandarin, and English), its distinct socio-legal context under the "one country, two systems" framework, and specific local cultural and value considerations, the model is built upon the DeepSeek architecture and systematically aligned with regional norms through a multifaceted full parameter fine-tuning process. It is further integrated with a retrieval-augmented generation (RAG) system to ensure timely and factually grounded information access. The core contribution lies in the design and implementation of a comprehensive, region-specific AI alignment and safety framework, demonstrated through two key achievements: 1) The successful development of HKGAI-V1 itself - which outper-forms general-purpose models in handling Hong Kong-specific culturally sensitive queries, and embodies a "governance-embedded" approach to digital sovereignty - empowers Hong Kong to exercise control over AI applications in critical sectors including public services, legal systems, and edu-cation. 2) The development of the proprietary Adversarial HK Value Benchmark, a rigorous tool for evaluating model alignment with local ethical and legal stand-ards under challenging conditions. By documenting these achievements, the paper provides not only a technological artifact but also a replicable blueprint for developing advanced, regionally focused AI systems deeply rooted in their local identities.
♻ ☆ Expert-level validation of AI-generated medical text with scalable language models
With the growing use of language models (LMs) in clinical environments, there is an immediate need to evaluate the accuracy and safety of LM-generated medical text. Currently, such evaluation relies solely on manual physician review. However, detecting errors in LM-generated text is challenging because 1) manual review is costly and 2) expert-composed reference outputs are often unavailable in real-world settings. While the "LM-as-judge" paradigm (a LM evaluating another LM) offers scalable evaluation, even frontier LMs can miss subtle but clinically significant errors. To address these challenges, we propose MedVAL, a self-supervised framework that leverages synthetic data to train evaluator LMs to assess whether LM-generated medical outputs are factually consistent with inputs, without requiring physician labels or reference outputs. To evaluate LM performance, we introduce MedVAL-Bench, a dataset containing 840 outputs annotated by physicians, following a physician-defined taxonomy of risk levels and error categories. Across 6 diverse medical tasks and 10 state-of-the-art LMs spanning open-source, proprietary, and medically adapted models, MedVAL fine-tuning significantly improves (p < 0.001) alignment with physicians on both seen and unseen tasks, increasing average F1 scores from 66% to 83%, with per-sample safety classification scores up to 86%. MedVAL improves the performance of even the best-performing proprietary LM (GPT-4o) by 8%. To support a scalable, risk-aware pathway towards clinical integration, we open-source the 1) codebase (https://github.com/StanfordMIMI/MedVAL), 2) MedVAL-Bench (https://huggingface.co/datasets/stanfordmimi/MedVAL-Bench), and 3) MedVAL-4B (https://huggingface.co/stanfordmimi/MedVAL-4B), the best-performing open-source LM. Our research provides the first evidence of LMs approaching expert-level validation ability for medical text.
♻ ☆ Leveraging Large Language Models for Multi-Class and Multi-Label Detection of Drug Use and Overdose Symptoms on Social Media
Drug overdose remains a critical global health issue, often driven by misuse of opioids, painkillers, and psychiatric medications. Traditional research methods face limitations, whereas social media offers real-time insights into self-reported substance use and overdose symptoms. This study proposes an AI-driven NLP framework trained on annotated social media data to detect commonly used drugs and associated overdose symptoms. Using a hybrid annotation strategy with LLMs and human annotators, we applied traditional ML models, neural networks, and advanced transformer-based models. Our framework achieved 98% accuracy in multi-class and 97% in multi-label classification, outperforming baseline models by up to 8%. These findings highlight the potential of AI for supporting public health surveillance and personalized intervention strategies.
♻ ☆ Roll the dice & look before you leap: Going beyond the creative limits of next-token prediction ICML 2025
We design a suite of minimal algorithmic tasks that are a loose abstraction of open-ended real-world tasks. This allows us to cleanly and controllably quantify the creative limits of the present-day language model. Much like real-world tasks that require a creative, far-sighted leap of thought, our tasks require an implicit, open-ended stochastic planning step that either (a) discovers new connections in an abstract knowledge graph (like in wordplay, drawing analogies, or research) or (b) constructs new patterns (like in designing math problems or new proteins). In these tasks, we empirically and conceptually argue how next-token learning is myopic; multi-token approaches, namely teacherless training and diffusion models, comparatively excel in producing diverse and original output. Secondly, to elicit randomness without hurting coherence, we find that injecting noise at the input layer (dubbed seed-conditioning) works surprisingly as well as (and in some conditions, better than) temperature sampling from the output layer. Thus, our work offers a principled, minimal test-bed for analyzing open-ended creative skills, and offers new arguments for going beyond next-token learning and temperature sampling. We make part of the code available under https://github.com/chenwu98/algorithmic-creativity
comment: ICML 2025 (oral)
♻ ☆ Beyond classical and contemporary models: a transformative AI framework for student dropout prediction in distance learning using RAG, Prompt engineering, and Cross-modal fusion
Student dropout in distance learning remains a critical challenge, with profound societal and economic consequences. While classical machine learning models leverage structured socio-demographic and behavioral data, they often fail to capture the nuanced emotional and contextual factors embedded in unstructured student interactions. This paper introduces a transformative AI framework that redefines dropout prediction through three synergistic innovations: Retrieval-Augmented Generation (RAG) for domain-specific sentiment analysis, prompt engineering to decode academic stressors,and cross-modal attention fusion to dynamically align textual, behavioral, and socio-demographic insights. By grounding sentiment analysis in a curated knowledge base of pedagogical content, our RAG-enhanced BERT model interprets student comments with unprecedented contextual relevance, while optimized prompts isolate indicators of academic distress (e.g., "isolation," "workload anxiety"). A cross-modal attention layer then fuses these insights with temporal engagement patterns, creating holistic risk pro-files. Evaluated on a longitudinal dataset of 4 423 students, the framework achieves 89% accuracy and an F1-score of 0.88, outperforming conventional models by 7% and reducing false negatives by 21%. Beyond prediction, the system generates interpretable interventions by retrieving contextually aligned strategies (e.g., mentorship programs for isolated learners). This work bridges the gap between predictive analytics and actionable pedagogy, offering a scalable solution to mitigate dropout risks in global education systems
comment: 13 pages, 8 figures, 1 Algorithms, 17th International Conference on Education and New Learning Technologies,: 30 June-2 July, 2025 Location: Palma, Spain
♻ ☆ EVOLvE: Evaluating and Optimizing LLMs For In-Context Exploration ICML 2025
Despite their success in many domains, large language models (LLMs) remain under-studied in scenarios requiring optimal decision-making under uncertainty. This is crucial as many real-world applications, ranging from personalized recommendations to healthcare interventions, demand that LLMs not only predict but also actively learn to make optimal decisions through exploration. In this work, we measure LLMs' (in)ability to make optimal decisions in bandits, a state-less reinforcement learning setting relevant to many applications. We develop a comprehensive suite of environments, including both context-free and contextual bandits with varying task difficulties, to benchmark LLMs' performance. Motivated by the existence of optimal exploration algorithms, we propose efficient ways to integrate this algorithmic knowledge into LLMs: by providing explicit algorithm-guided support during inference; and through algorithm distillation via in-context demonstrations and fine-tuning, using synthetic data generated from these algorithms. Impressively, these techniques allow us to achieve superior exploration performance with smaller models, surpassing larger models on various tasks. We conducted an extensive ablation study to shed light on various factors, such as task difficulty and data representation, that influence the efficiency of LLM exploration. Additionally, we conduct a rigorous analysis of the LLM's exploration efficiency using the concept of regret, linking its ability to explore to the model size and underlying algorithm.
comment: 28 pages. Published at ICML 2025
♻ ☆ ContextCache: Context-Aware Semantic Cache for Multi-Turn Queries in Large Language Models
Semantic caching significantly reduces computational costs and improves efficiency by storing and reusing large language model (LLM) responses. However, existing systems rely primarily on matching individual queries, lacking awareness of multi-turn dialogue contexts, which leads to incorrect cache hits when similar queries appear in different conversational settings. This demonstration introduces ContextCache, a context-aware semantic caching system for multi-turn dialogues. ContextCache employs a two-stage retrieval architecture that first executes vector-based retrieval on the current query to identify potential matches and then integrates current and historical dialogue representations through self-attention mechanisms for precise contextual matching. Evaluation of real-world conversations shows that ContextCache improves precision and recall compared to existing methods. Additionally, cached responses exhibit approximately 10 times lower latency than direct LLM invocation, enabling significant computational cost reductions for LLM conversational applications.
♻ ☆ LEXam: Benchmarking Legal Reasoning on 340 Law Exams
Long-form legal reasoning remains a key challenge for large language models (LLMs) in spite of recent advances in test-time scaling. We introduce LEXam, a novel benchmark derived from 340 law exams spanning 116 law school courses across a range of subjects and degree levels. The dataset comprises 4,886 law exam questions in English and German, including 2,841 long-form, open-ended questions and 2,045 multiple-choice questions. Besides reference answers, the open questions are also accompanied by explicit guidance outlining the expected legal reasoning approach such as issue spotting, rule recall, or rule application. Our evaluation on both open-ended and multiple-choice questions present significant challenges for current LLMs; in particular, they notably struggle with open questions that require structured, multi-step legal reasoning. Moreover, our results underscore the effectiveness of the dataset in differentiating between models with varying capabilities. Adopting an LLM-as-a-Judge paradigm with rigorous human expert validation, we demonstrate how model-generated reasoning steps can be evaluated consistently and accurately. Our evaluation setup provides a scalable method to assess legal reasoning quality beyond simple accuracy metrics. Project page: https://lexam-benchmark.github.io/
♻ ☆ Bias Beyond English: Evaluating Social Bias and Debiasing Methods in a Low-Resource Setting
Social bias in language models can potentially exacerbate social inequalities. Despite it having garnered wide attention, most research focuses on English data. In a low-resource scenario, the models often perform worse due to insufficient training data. This study aims to leverage high-resource language corpora to evaluate bias and experiment with debiasing methods in low-resource languages. We evaluated the performance of recent multilingual models in five languages: English, Chinese, Russian, Indonesian and Thai, and analyzed four bias dimensions: gender, religion, nationality, and race-color. By constructing multilingual bias evaluation datasets, this study allows fair comparisons between models across languages. We have further investigated three debiasing methods-CDA, Dropout, SenDeb-and demonstrated that debiasing methods from high-resource languages can be effectively transferred to low-resource ones, providing actionable insights for fairness research in multilingual NLP.
♻ ☆ B-cos LM: Efficiently Transforming Pre-trained Language Models for Improved Explainability
Post-hoc explanation methods for black-box models often struggle with faithfulness and human interpretability due to the lack of explainability in current neural architectures. Meanwhile, B-cos networks have been introduced to improve model explainability by proposing an architecture that removes bias terms and promotes input-weight alignment. Although B-cos networks have shown success in building explainable systems, their application has so far been limited to computer vision models and their associated training pipelines. In this work, we introduce B-cos LMs, i.e., B-cos language models (LMs) empowered for natural language processing (NLP) tasks. Our approach directly transforms pre-trained language models into B-cos LMs by combining B-cos conversion and task fine-tuning, improving efficiency compared to previous methods. Our automatic and human evaluation results demonstrate that B-cos LMs produce more faithful and human interpretable explanations than post-hoc methods, while maintaining task performance comparable to conventional fine-tuning. Our in-depth analysis explores how B-cos LMs differ from conventionally fine-tuned models in their learning processes and explanation patterns. Finally, we are also the first to explore the transformation of decoder-only models to B-cos LMs for generation tasks.
♻ ☆ The distribution of syntactic dependency distances
The syntactic structure of a sentence can be represented as a graph, where vertices are words and edges indicate syntactic dependencies between them. In this setting, the distance between two linked words is defined as the difference between their positions. Here we wish to contribute to the characterization of the actual distribution of syntactic dependency distances, which has previously been argued to follow a power-law distribution. Here we propose a new model with two exponential regimes in which the probability decay is allowed to change after a break-point. This transition could mirror the transition from the processing of word chunks to higher-level structures. We find that a two-regime model - where the first regime follows either an exponential or a power-law decay - is the most likely one in all 20 languages we considered, independently of sentence length and annotation style. Moreover, the break-point exhibits low variation across languages and averages values of 4-5 words, suggesting that the amount of words that can be simultaneously processed abstracts from the specific language to a high degree. The probability decay slows down after the breakpoint, consistently with a universal chunk-and-pass mechanism. Finally, we give an account of the relation between the best estimated model and the closeness of syntactic dependencies as function of sentence length, according to a recently introduced optimality score.
comment: minor corrections; in press in Glottometrics
♻ ☆ Trinity-RFT: A General-Purpose and Unified Framework for Reinforcement Fine-Tuning of Large Language Models
Trinity-RFT is a general-purpose, unified and easy-to-use framework designed for reinforcement fine-tuning (RFT) of large language models. It is built with a modular and decoupled design, consisting of (1) an RFT-core that unifies and generalizes synchronous/asynchronous, on-policy/off-policy, and online/offline modes of RFT; (2) seamless integration for agent-environment interaction with high efficiency and robustness; and (3) systematic data pipelines optimized for RFT. Trinity-RFT can be easily adapted for diverse application scenarios, and serves as a unified platform for development and research of advanced reinforcement learning paradigms at both macroscopic and microscopic levels. This technical report outlines the vision, features, design and implementations of Trinity-RFT, accompanied by extensive examples, applications and experiments that demonstrate its functionalities and user-friendliness.
comment: This technical report will be continuously updated as the codebase evolves. GitHub: https://github.com/modelscope/Trinity-RFT
♻ ☆ Mechanistic Indicators of Understanding in Large Language Models
Recent findings in mechanistic interpretability (MI), the field probing the inner workings of Large Language Models (LLMs), challenge the view that these models rely solely on superficial statistics. We offer an accessible synthesis of these findings that doubles as an introduction to MI while integrating these findings within a novel theoretical framework for thinking about machine understanding. We argue that LLMs develop internal structures that are functionally analogous to the kind of understanding that consists in seeing connections. To sharpen this idea, we propose a three-tiered conception of understanding. First, conceptual understanding emerges when a model forms "features" as directions in latent space, learning the connections between diverse manifestations of something. Second, state-of-the-world understanding emerges when a model learns contingent factual connections between features and dynamically tracks changes in the world. Third, principled understanding emerges when a model ceases to rely on a collection of memorized facts and discovers a "circuit" connecting these facts. However, these forms of understanding remain radically different from human understanding, as the phenomenon of "parallel mechanisms" shows. We conclude that the debate should move beyond the yes-or-no question of whether LLMs understand to investigate how their strange minds work and forge conceptions that fit them.
comment: 32 pages
♻ ☆ A Noise-Robust Turn-Taking System for Real-World Dialogue Robots: A Field Experiment IROS 2025
Turn-taking is a crucial aspect of human-robot interaction, directly influencing conversational fluidity and user engagement. While previous research has explored turn-taking models in controlled environments, their robustness in real-world settings remains underexplored. In this study, we propose a noise-robust voice activity projection (VAP) model, based on a Transformer architecture, to enhance real-time turn-taking in dialogue robots. To evaluate the effectiveness of the proposed system, we conducted a field experiment in a shopping mall, comparing the VAP system with a conventional cloud-based speech recognition system. Our analysis covered both subjective user evaluations and objective behavioral analysis. The results showed that the proposed system significantly reduced response latency, leading to a more natural conversation where both the robot and users responded faster. The subjective evaluations suggested that faster responses contribute to a better interaction experience.
comment: This paper has been accepted for presentation at IEEE/RSJ International Conference on Intelligent Robots and Systems 2025 (IROS 2025) and represents the author's version of the work
♻ ☆ Barriers in Integrating Medical Visual Question Answering into Radiology Workflows: A Scoping Review and Clinicians' Insights
Medical Visual Question Answering (MedVQA) is a promising tool to assist radiologists by automating medical image interpretation through question answering. Despite advances in models and datasets, MedVQA's integration into clinical workflows remains limited. This study systematically reviews 68 publications (2018-2024) and surveys 50 clinicians from India and Thailand to examine MedVQA's practical utility, challenges, and gaps. Following the Arksey and O'Malley scoping review framework, we used a two-pronged approach: (1) reviewing studies to identify key concepts, advancements, and research gaps in radiology workflows, and (2) surveying clinicians to capture their perspectives on MedVQA's clinical relevance. Our review reveals that nearly 60% of QA pairs are non-diagnostic and lack clinical relevance. Most datasets and models do not support multi-view, multi-resolution imaging, EHR integration, or domain knowledge, features essential for clinical diagnosis. Furthermore, there is a clear mismatch between current evaluation metrics and clinical needs. The clinician survey confirms this disconnect: only 29.8% consider MedVQA systems highly useful. Key concerns include the absence of patient history or domain knowledge (87.2%), preference for manually curated datasets (51.1%), and the need for multi-view image support (78.7%). Additionally, 66% favor models focused on specific anatomical regions, and 89.4% prefer dialogue-based interactive systems. While MedVQA shows strong potential, challenges such as limited multimodal analysis, lack of patient context, and misaligned evaluation approaches must be addressed for effective clinical integration.
comment: 29 pages, 5 figures (1 in supplementary), 3 tables (1 in main text, 2 in supplementary). Scoping review and clinician survey
♻ ☆ DiaTool-DPO: Multi-Turn Direct Preference Optimization for Tool-Augmented Large Language Models SIGDIAL 2025
Tool-Augmented Larage Language Models (TA-LLMs) have shown promise in real-world applications, but face challenges in handling incomplete queries and out-of-scope requests. While existing approaches rely mainly on Supervised Fine-Tuning with expert trajectories, we propose DiaTool-DPO, a novel method that enhances TA-LLM's dialogue capabilities through Direct Preference Optimization. We model TA-LLM interactions as a Markov Decision Process with 5 distinct dialogue states and categorize user queries into 3 types based on their state transition trajectories. We automatically construct paired trajectory datasets of correct and incorrect dialogue flows and introduce a specialized objective loss for dialogue control. Our comprehensive evaluation demonstrates that DiaTool-DPO approaches GPT-4o's performance (94.8% in information gathering, 91% in tool call rejection) with substantial improvements over baseline (44% and 9.6% respectively) while maintaining core functionality. Our approach opens new possibilities for developing TA-LLMs that can handle diverse real-world scenarios without requiring additional expert demonstrations or human labeling.
comment: Accepted to SIGDIAL 2025
♻ ☆ A Comprehensive Survey of Direct Preference Optimization: Datasets, Theories, Variants, and Applications
With the rapid advancement of large language models (LLMs), aligning policy models with human preferences has become increasingly critical. Direct Preference Optimization (DPO) has emerged as a promising approach for alignment, acting as an RL-free alternative to Reinforcement Learning from Human Feedback (RLHF). Despite DPO's various advancements and inherent limitations, an in-depth review of these aspects is currently lacking in the literature. In this work, we present a comprehensive review of the challenges and opportunities in DPO, covering theoretical analyses, variants, relevant preference datasets, and applications. Specifically, we categorize recent studies on DPO based on key research questions to provide a thorough understanding of DPO's current landscape. Additionally, we propose several future research directions to offer insights on model alignment for the research community. An updated collection of relevant papers can be found on https://github.com/Mr-Loevan/DPO-Survey.
comment: 45 pages, 12 Figures. Project page: https://github.com/Mr-Loevan/DPO-Survey
♻ ☆ Structuring Radiology Reports: Challenging LLMs with Lightweight Models
Radiology reports are critical for clinical decision-making but often lack a standardized format, limiting both human interpretability and machine learning (ML) applications. While large language models (LLMs) have shown strong capabilities in reformatting clinical text, their high computational requirements, lack of transparency, and data privacy concerns hinder practical deployment. To address these challenges, we explore lightweight encoder-decoder models (<300M parameters)-specifically T5 and BERT2BERT-for structuring radiology reports from the MIMIC-CXR and CheXpert Plus datasets. We benchmark these models against eight open-source LLMs (1B-70B), adapted using prefix prompting, in-context learning (ICL), and low-rank adaptation (LoRA) finetuning. Our best-performing lightweight model outperforms all LLMs adapted using prompt-based techniques on a human-annotated test set. While some LoRA-finetuned LLMs achieve modest gains over the lightweight model on the Findings section (BLEU 6.4%, ROUGE-L 4.8%, BERTScore 3.6%, F1-RadGraph 1.1%, GREEN 3.6%, and F1-SRR-BERT 4.3%), these improvements come at the cost of substantially greater computational resources. For example, LLaMA-3-70B incurred more than 400 times the inference time, cost, and carbon emissions compared to the lightweight model. These results underscore the potential of lightweight, task-specific models as sustainable and privacy-preserving solutions for structuring clinical text in resource-constrained healthcare settings.
♻ ☆ Following the Clues: Experiments on Person Re-ID using Cross-Modal Intelligence SC 2025
The collection and release of street-level recordings as Open Data play a vital role in advancing autonomous driving systems and AI research. However, these datasets pose significant privacy risks, particularly for pedestrians, due to the presence of Personally Identifiable Information (PII) that extends beyond biometric traits such as faces. In this paper, we present cRID, a novel cross-modal framework combining Large Vision-Language Models, Graph Attention Networks, and representation learning to detect textual describable clues of PII and enhance person re-identification (Re-ID). Our approach focuses on identifying and leveraging interpretable features, enabling the detection of semantically meaningful PII beyond low-level appearance cues. We conduct a systematic evaluation of PII presence in person image datasets. Our experiments show improved performance in practical cross-dataset Re-ID scenarios, notably from Market-1501 to CUHK03-np (detected), highlighting the framework's practical utility. Code is available at https://github.com/RAufschlaeger/cRID.
comment: accepted for publication at the 2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC 2025), taking place during November 18-21, 2025 in Gold Coast, Australia
♻ ☆ Political Bias in LLMs: Unaligned Moral Values in Agent-centric Simulations
Contemporary research in social sciences increasingly utilizes state-of-the-art generative language models to annotate or generate content. While these models achieve benchmark-leading performance on common language tasks, their application to novel out-of-domain tasks remains insufficiently explored. To address this gap, we investigate how personalized language models align with human responses on the Moral Foundation Theory Questionnaire. We adapt open-source generative language models to different political personas and repeatedly survey these models to generate synthetic data sets where model-persona combinations define our sub-populations. Our analysis reveals that models produce inconsistent results across multiple repetitions, yielding high response variance. Furthermore, the alignment between synthetic data and corresponding human data from psychological studies shows a weak correlation, with conservative persona-prompted models particularly failing to align with actual conservative populations. These results suggest that language models struggle to coherently represent ideologies through in-context prompting due to their alignment process. Thus, using language models to simulate social interactions requires measurable improvements in in-context optimization or parameter manipulation to align with psychological and sociological stereotypes properly.
comment: 14 pages, 2 tables
♻ ☆ IPAD: Inverse Prompt for AI Detection -- A Robust and Explainable LLM-Generated Text Detector
Large Language Models (LLMs) have attained human-level fluency in text generation, which complicates the distinction between human-written and LLM-generated texts. This increases the risk of misuse and highlights the need for reliable detectors. Yet, existing detectors exhibit poor robustness on out-of-distribution (OOD) data and attacked data, which is critical for real-world scenarios. Also, they struggle to provide interpretable evidence to support their decisions, thus undermining the reliability. In light of these challenges, we propose IPAD (Inverse Prompt for AI Detection), a novel framework consisting of a Prompt Inverter that identifies predicted prompts that could have generated the input text, and two Distinguishers that examine the probability that the input texts align with the predicted prompts. Empirical evaluations demonstrate that IPAD outperforms the strongest baselines by 9.05% (Average Recall) on in-distribution data, 12.93% (AUROC) on out-of-distribution (OOD) data, and 5.48% (AUROC) on attacked data. IPAD also performs robustly on structured datasets. Furthermore, an interpretability assessment is conducted to illustrate that IPAD enhances the AI detection trustworthiness by allowing users to directly examine the decision-making evidence, which provides interpretable support for its state-of-the-art detection results.
♻ ☆ Scientists' First Exam: Probing Cognitive Abilities of MLLM via Perception, Understanding, and Reasoning
Scientific discoveries increasingly rely on complex multimodal reasoning based on information-intensive scientific data and domain-specific expertise. Empowered by expert-level scientific benchmarks, scientific Multimodal Large Language Models (MLLMs) hold the potential to significantly enhance this discovery process in realistic workflows. However, current scientific benchmarks mostly focus on evaluating the knowledge understanding capabilities of MLLMs, leading to an inadequate assessment of their perception and reasoning abilities. To address this gap, we present the Scientists' First Exam (SFE) benchmark, designed to evaluate the scientific cognitive capacities of MLLMs through three interconnected levels: scientific signal perception, scientific attribute understanding, scientific comparative reasoning. Specifically, SFE comprises 830 expert-verified VQA pairs across three question types, spanning 66 multimodal tasks across five high-value disciplines. Extensive experiments reveal that current state-of-the-art GPT-o3 and InternVL-3 achieve only 34.08% and 26.52% on SFE, highlighting significant room for MLLMs to improve in scientific realms. We hope the insights obtained in SFE will facilitate further developments in AI-enhanced scientific discoveries.
comment: 82 pages
♻ ☆ SpatialViz-Bench: Automatically Generated Spatial Visualization Reasoning Tasks for MLLMs
Humans can directly imagine and manipulate visual images in their minds, a capability known as spatial visualization. While multi-modal Large Language Models (MLLMs) support imagination-based reasoning, spatial visualization remains insufficiently evaluated, typically embedded within broader mathematical and logical assessments. Existing evaluations often rely on IQ tests or math competitions that may overlap with training data, compromising assessment reliability. To this end, we introduce SpatialViz-Bench, a comprehensive multi-modal benchmark for spatial visualization with 12 tasks across 4 sub-abilities, comprising 1,180 automatically generated problems. Our evaluation of 33 state-of-the-art MLLMs not only reveals wide performance variations and demonstrates the benchmark's strong discriminative power, but also uncovers counter-intuitive findings: models exhibit unexpected behaviors by showing difficulty perception that misaligns with human intuition, displaying dramatic 2D-to-3D performance cliffs, and defaulting to formula derivation despite spatial tasks requiring visualization alone. SpatialVizBench empirically demonstrates that state-of-the-art MLLMs continue to exhibit deficiencies in spatial visualization tasks, thereby addressing a significant lacuna in the field. The benchmark is publicly available.
♻ ☆ Teaching LLM to Reason: Reinforcement Learning from Algorithmic Problems without Code
Enhancing reasoning capabilities remains a central focus in the LLM reasearch community. A promising direction involves requiring models to simulate code execution step-by-step to derive outputs for given inputs. However, as code is often designed for large-scale systems, direct application leads to over-reliance on complex data structures and algorithms, even for simple cases, resulting in overfitting to algorithmic patterns rather than core reasoning structures. To address this, we propose TeaR, which aims at teaching LLMs to reason better. TeaR leverages careful data curation and reinforcement learning to guide models in discovering optimal reasoning paths through code-related tasks, thereby improving general reasoning abilities. We conduct extensive experiments using two base models and three long-CoT distillation models, with model sizes ranging from 1.5 billion to 32 billion parameters, and across 17 benchmarks spanning Math, Knowledge, Code, and Logical Reasoning. The results consistently show significant performance improvements. Notably, TeaR achieves a 35.9% improvement on Qwen2.5-7B and 5.9% on R1-Distilled-7B.
♻ ☆ Not all tokens are created equal: Perplexity Attention Weighted Networks for AI generated text detection
The rapid advancement in large language models (LLMs) has significantly enhanced their ability to generate coherent and contextually relevant text, raising concerns about the misuse of AI-generated content and making it critical to detect it. However, the task remains challenging, particularly in unseen domains or with unfamiliar LLMs. Leveraging LLM next-token distribution outputs offers a theoretically appealing approach for detection, as they encapsulate insights from the models' extensive pre-training on diverse corpora. Despite its promise, zero-shot methods that attempt to operationalize these outputs have met with limited success. We hypothesize that one of the problems is that they use the mean to aggregate next-token distribution metrics across tokens, when some tokens are naturally easier or harder to predict and should be weighted differently. Based on this idea, we propose the Perplexity Attention Weighted Network (PAWN), which uses the last hidden states of the LLM and positions to weight the sum of a series of features based on metrics from the next-token distribution across the sequence length. Although not zero-shot, our method allows us to cache the last hidden states and next-token distribution metrics on disk, greatly reducing the training resource requirements. PAWN shows competitive and even better performance in-distribution than the strongest baselines (fine-tuned LMs) with a fraction of their trainable parameters. Our model also generalizes better to unseen domains and source models, with smaller variability in the decision boundary across distribution shifts. It is also more robust to adversarial attacks, and if the backbone has multilingual capabilities, it presents decent generalization to languages not seen during supervised training, with LLaMA3-1B reaching a mean macro-averaged F1 score of 81.46% in cross-validation with nine languages.
♻ ☆ TReB: A Comprehensive Benchmark for Evaluating Table Reasoning Capabilities of Large Language Models
The majority of data in businesses and industries is stored in tables, databases, and data warehouses. Reasoning with table-structured data poses significant challenges for large language models (LLMs) due to its hidden semantics, inherent complexity, and structured nature. One of these challenges is lacking an effective evaluation benchmark fairly reflecting the performances of LLMs on broad table reasoning abilities. In this paper, we fill in this gap, presenting a comprehensive table reasoning evolution benchmark, TReB, which measures both shallow table understanding abilities and deep table reasoning abilities, a total of 26 sub-tasks. We construct a high quality dataset through an iterative data processing procedure. We create an evaluation framework to robustly measure table reasoning capabilities with three distinct inference modes, TCoT, PoT and ICoT. Further, we benchmark over 20 state-of-the-art LLMs using this frame work and prove its effectiveness. Experimental results reveal that existing LLMs still have significant room for improvement in addressing the complex and real world Table related tasks. Both the dataset and evaluation framework are publicly available, with the dataset hosted on huggingface.co/datasets/JT-LM/JIUTIAN-TReB and the framework on github.com/JT-LM/jiutian-treb.
comment: Benmark report v1.1
♻ ☆ PRIME: Large Language Model Personalization with Cognitive Memory and Thought Processes
Large language model (LLM) personalization aims to align model outputs with individuals' unique preferences and opinions. While recent efforts have implemented various personalization methods, a unified theoretical framework that can systematically understand the drivers of effective personalization is still lacking. In this work, we integrate the well-established cognitive dual-memory model into LLM personalization, by mirroring episodic memory to historical user engagements and semantic memory to long-term, evolving user beliefs. Specifically, we systematically investigate memory instantiations and introduce a unified framework, PRIME, using episodic and semantic memory mechanisms. We further augment PRIME with a novel personalized thinking capability inspired by the slow thinking strategy. Moreover, recognizing the absence of suitable benchmarks, we introduce a dataset using Change My View (CMV) from Reddit, specifically designed to evaluate long-context personalization. Extensive experiments validate PRIME's effectiveness across both long- and short-context scenarios. Further analysis confirms that PRIME effectively captures dynamic personalization beyond mere popularity biases.
♻ ☆ DeepGesture: A conversational gesture synthesis system based on emotions and semantics
Along with the explosion of large language models, improvements in speech synthesis, advancements in hardware, and the evolution of computer graphics, the current bottleneck in creating digital humans lies in generating character movements that correspond naturally to text or speech inputs. In this work, we present DeepGesture, a diffusion-based gesture synthesis framework for generating expressive co-speech gestures conditioned on multimodal signals - text, speech, emotion, and seed motion. Built upon the DiffuseStyleGesture model, DeepGesture introduces novel architectural enhancements that improve semantic alignment and emotional expressiveness in generated gestures. Specifically, we integrate fast text transcriptions as semantic conditioning and implement emotion-guided classifier-free diffusion to support controllable gesture generation across affective states. To visualize results, we implement a full rendering pipeline in Unity based on BVH output from the model. Evaluation on the ZeroEGGS dataset shows that DeepGesture produces gestures with improved human-likeness and contextual appropriateness. Our system supports interpolation between emotional states and demonstrates generalization to out-of-distribution speech, including synthetic voices - marking a step forward toward fully multimodal, emotionally aware digital humans. Project page: https://deepgesture.github.io
comment: Project page: https://deepgesture.github.io
♻ ☆ EVALOOP: Assessing LLM Robustness in Programming from a Self-consistency Perspective
Assessing the programming capabilities of Large Language Models (LLMs) is crucial for their effective use in software engineering. Current evaluations, however, predominantly measure the accuracy of generated code on static benchmarks, neglecting the critical aspect of model robustness during programming tasks. While adversarial attacks offer insights on model robustness, their effectiveness is limited and evaluation could be constrained. Current adversarial attack methods for robustness evaluation yield inconsistent results, struggling to provide a unified evaluation across different LLMs. We introduce EVALOOP, a novel assessment framework that evaluate the robustness from a self-consistency perspective, i.e., leveraging the natural duality inherent in popular software engineering tasks, e.g., code generation and code summarization. EVALOOP initiates a self-contained feedback loop: an LLM generates output (e.g., code) from an input (e.g., natural language specification), and then use the generated output as the input to produce a new output (e.g., summarizes that code into a new specification). EVALOOP repeats the process to assess the effectiveness of EVALOOP in each loop. This cyclical strategy intrinsically evaluates robustness without rely on any external attack setups, providing a unified metric to evaluate LLMs' robustness in programming. We evaluate 16 prominent LLMs (e.g., GPT-4.1, O4-mini) on EVALOOP and found that EVALOOP typically induces a 5.01%-19.31% absolute drop in pass@1 performance within ten loops. Intriguingly, robustness does not always align with initial performance (i.e., one-time query); for instance, GPT-3.5-Turbo, despite superior initial code generation compared to DeepSeek-V2, demonstrated lower robustness over repeated evaluation loop.
comment: 20 pages, 11 figures
♻ ☆ Qorgau: Evaluating LLM Safety in Kazakh-Russian Bilingual Contexts
Large language models (LLMs) are known to have the potential to generate harmful content, posing risks to users. While significant progress has been made in developing taxonomies for LLM risks and safety evaluation prompts, most studies have focused on monolingual contexts, primarily in English. However, language- and region-specific risks in bilingual contexts are often overlooked, and core findings can diverge from those in monolingual settings. In this paper, we introduce Qorgau, a novel dataset specifically designed for safety evaluation in Kazakh and Russian, reflecting the unique bilingual context in Kazakhstan, where both Kazakh (a low-resource language) and Russian (a high-resource language) are spoken. Experiments with both multilingual and language-specific LLMs reveal notable differences in safety performance, emphasizing the need for tailored, region-specific datasets to ensure the responsible and safe deployment of LLMs in countries like Kazakhstan. Warning: this paper contains example data that may be offensive, harmful, or biased.
♻ ☆ ACEBench: Who Wins the Match Point in Tool Usage?
Large Language Models (LLMs) have demonstrated significant potential in decision-making and reasoning, particularly when integrated with various tools to effectively solve complex problems. However, existing benchmarks for evaluating LLMs' tool usage face several limitations: (1) limited evaluation scenarios, often lacking assessments in real multi-turn dialogue contexts; (2) narrow evaluation dimensions, with insufficient detailed assessments of how LLMs use tools; and (3) reliance on LLMs or real API executions for evaluation, which introduces significant overhead. To address these challenges, we introduce ACEBench, a comprehensive benchmark for assessing tool usage in LLMs. ACEBench categorizes data into three primary types based on evaluation methodology: Normal, Special, and Agent. "Normal" evaluates tool usage in basic scenarios; "Special" evaluates tool usage in situations with ambiguous or incomplete instructions; "Agent" evaluates tool usage through multi-agent interactions to simulate real-world, multi-turn dialogues. We conducted extensive experiments using ACEBench, analyzing various LLMs in-depth and providing a more granular examination of error causes across different data types.
♻ ☆ PyVision: Agentic Vision with Dynamic Tooling
LLMs are increasingly deployed as agents, systems capable of planning, reasoning, and dynamically calling external tools. However, in visual reasoning, prior approaches largely remain limited by predefined workflows and static toolsets. In this report, we present PyVision, an interactive, multi-turn framework that enables MLLMs to autonomously generate, execute, and refine Python-based tools tailored to the task at hand, unlocking flexible and interpretable problem-solving. We develop a taxonomy of the tools created by PyVision and analyze their usage across a diverse set of benchmarks. Quantitatively, PyVision achieves consistent performance gains, boosting GPT-4.1 by +7.8% on V* and Claude-4.0-Sonnet by +31.1% on VLMsAreBlind-mini. These results point to a broader shift: dynamic tooling allows models not just to use tools, but to invent them, advancing toward more agentic visual reasoning.
comment: 26 Pages, 10 Figures, Technical report
♻ ☆ Fourier Position Embedding: Enhancing Attention's Periodic Extension for Length Generalization ICML 2025
Extending the context length of Language Models (LMs) by improving Rotary Position Embedding (RoPE) has become a trend. While prior works mainly address RoPE's limitations within attention, this paper uncovers the adverse effects on length generalization from nearly all parts of LMs. Using Discrete Signal Processing theory, we show that RoPE enables periodic attention by implicitly achieving Non-Uniform Discrete Fourier Transform. However, this periodicity is undermined by the spectrum damage caused by: 1) linear layers and activation functions; 2) insufficiently trained frequency components brought by time-domain truncation. Building on our observations, we propose Fourier Position Embedding (FoPE), which enhances attention's frequency-domain properties to improve both its periodic extension and length generalization. FoPE constructs \textit{Fourier Series} and zero-outs the destructive frequency components, increasing model robustness against the spectrum damage. Experiments across various model scales and benchmarks show that, within varying context windows, FoPE maintains a more stable performance compared to other baselines. Several analyses and ablations bring further support to our method and theoretical modeling.
comment: Accepted to ICML 2025
♻ ☆ Intuitive Fine-Tuning: Towards Simplifying Alignment into a Single Process ACL 2025
Supervised Fine-Tuning (SFT) and Preference Optimization (PO) are key processes for aligning Language Models (LMs) with human preferences post pre-training. While SFT excels in efficiency and PO in effectiveness, they are often combined sequentially without integrating their optimization objectives. This approach ignores the opportunities to bridge their paradigm gap and take the strengths from both. In this paper, we interpret SFT and PO with two sub-processes -- Preference Estimation and Transition Optimization -- defined at token level within the Markov Decision Process (MDP). This modeling shows that SFT is only a special case of PO with inferior estimation and optimization. PO estimates the model's preference by its entire generation, while SFT only scores model's subsequent predicted tokens based on prior tokens from ground truth answer. These priors deviates from model's distribution, hindering the preference estimation and transition optimization. Building on this view, we introduce Intuitive Fine-Tuning (IFT) to integrate SFT and PO into a single process. Through a temporal residual connection, IFT brings better estimation and optimization by capturing LMs' intuitive sense of its entire answers. But it solely relies on a single policy and the same volume of non-preference-labeled data as SFT. Our experiments show that IFT performs comparably or even superiorly to SFT and some typical PO methods across several tasks, particularly those require generation, reasoning, and fact-following abilities. An explainable Frozen Lake game further validates the effectiveness of IFT for getting competitive policy.
comment: Accepted to ACL 2025, Oral & Panel Discussion
♻ ☆ Scalable MatMul-free Language Modeling
Large Language Models (LLMs) have fundamentally altered how we approach scaling in machine learning. However, these models pose substantial computational and memory challenges, primarily due to the reliance on matrix multiplication (MatMul) within their attention and feed-forward (FFN) layers. We demonstrate that MatMul operations can be eliminated from LLMs while maintaining strong performance, even at billion-parameter scales. Our MatMul-free models, tested on models up to 2.7B parameters, are comparable to state-of-the-art pre-trained Transformers, and the performance gap narrows as model size increases. Our approach yields significant memory savings: a GPU-efficient implementation reduces memory consumption by up to 61\% during training and over 10$\times$ during inference. When adapted for a multi-chip neuromorphic system, the model leverages asynchronous processing to achieve 4$\times$ higher throughput with 10$\times$ less energy than edge GPUs. %and 77$\times$ less energy than server-class GPUs, demonstrating superior scaling. These findings demonstrate a path toward dramatically simplified yet effective LLMs, advancing them toward brain-like efficiency and heralding a new generation of lightweight, high-performance language models. Our code implementation is available at https://github. com/ridgerchu/matmulfreellm.
♻ ☆ CV-Probes: Studying the interplay of lexical and world knowledge in visually grounded verb understanding
How do vision-language (VL) transformer models ground verb phrases and do they integrate contextual and world knowledge in this process? We introduce the CV-Probes dataset, containing image-caption pairs involving verb phrases that require both social knowledge and visual context to interpret (e.g., "beg"), as well as pairs involving verb phrases that can be grounded based on information directly available in the image (e.g., "sit"). We show that VL models struggle to ground VPs that are strongly context-dependent. Further analysis using explainable AI techniques shows that such models may not pay sufficient attention to the verb token in the captions. Our results suggest a need for improved methodologies in VL model training and evaluation. The code and dataset will be available https://github.com/ivana-13/CV-Probes.
comment: 9 pages, 2 figure, 6 tables, CogSci conference 2025
♻ ☆ InstCache: A Predictive Cache for LLM Serving
The revolutionary capabilities of Large Language Models (LLMs) are attracting rapidly growing popularity and leading to soaring user requests to inference serving systems. Caching techniques, which leverage data reuse to reduce computation, offer opportunities to optimize the performance of LLM inference engines. On the one hand, the low-level key-value (KV) cache working at the token level is widely adopted, albeit it incurs significant overhead as request volume grows. On the other hand, instruction-level caching, which stores full instruction-response pairs, is expected to play an increasingly crucial role. However, the high variability in the content and length of instructions make it rare for identical instructions to recur within a short time window, presenting challenges for effective caching instruction-response pairs. To address this challenge, we propose InstCache, a predictive caching mechanism for LLM serving systems. Leveraging the capability of LLMs, we can effectively reorder the representation space of instruction texts and develop a sufficient level of spatial locality. Such spatial locality enables us to predict potential instructions located in a compact region in the space, resulting in an effective caching system at runtime. Experimental results demonstrate that InstCache achieves a 2.3x higher hit rate compared to the upper bound of traditional caching mechanisms on WildChat dataset and reduces the time per output token of vLLM by up to 42.0% and 50.0% on LMSys and Moss datasets, respectively.
♻ ☆ BIS Reasoning 1.0: The First Large-Scale Japanese Benchmark for Belief-Inconsistent Syllogistic Reasoning
We present BIS Reasoning 1.0, the first large-scale Japanese dataset of syllogistic reasoning problems explicitly designed to evaluate belief-inconsistent reasoning in large language models (LLMs). Unlike prior datasets such as NeuBAROCO and JFLD, which focus on general or belief-aligned reasoning, BIS Reasoning 1.0 introduces logically valid yet belief-inconsistent syllogisms to uncover reasoning biases in LLMs trained on human-aligned corpora. We benchmark state-of-the-art models - including GPT models, Claude models, and leading Japanese LLMs - revealing significant variance in performance, with GPT-4o achieving 79.54% accuracy. Our analysis identifies critical weaknesses in current LLMs when handling logically valid but belief-conflicting inputs. These findings have important implications for deploying LLMs in high-stakes domains such as law, healthcare, and scientific literature, where truth must override intuitive belief to ensure integrity and safety.
comment: This version includes minor typo corrections in the example image
♻ ☆ REINFORCE++: An Efficient RLHF Algorithm with Robustness to Both Prompt and Reward Models
Reinforcement Learning from Human Feedback (RLHF) plays a crucial role in aligning large language models (LLMs) with human values and preferences. While state-of-the-art applications like ChatGPT/GPT-4 commonly employ Proximal Policy Optimization (PPO), the inclusion of a critic network introduces significant computational overhead. REINFORCE-based methods, such as REINFORCE Leave One-Out (RLOO), ReMax, and Group Relative Policy Optimization (GRPO), address this limitation by eliminating the critic network. However, these approaches face challenges in accurate advantage estimation. Specifically, they estimate advantages independently for responses to each prompt, which can lead to overfitting on simpler prompts and vulnerability to reward hacking. To address these challenges, we introduce REINFORCE++, a novel approach that removes the critic model while using the normalized reward of a batch as the baseline. Our empirical evaluation demonstrates that REINFORCE++ exhibits robust performance across various reward models without requiring prompt set truncation. Furthermore, it achieves superior generalization in both RLHF and long chain-of-thought (CoT) settings compared to existing REINFORCE-based methods. The implementation is available at https://github.com/OpenRLHF/OpenRLHF.
comment: add proof
♻ ☆ A General Framework for Inference-time Scaling and Steering of Diffusion Models
Diffusion models produce impressive results in modalities ranging from images and video to protein design and text. However, generating samples with user-specified properties remains a challenge. Recent research proposes fine-tuning models to maximize rewards that capture desired properties, but these methods require expensive training and are prone to mode collapse. In this work, we present Feynman-Kac (FK) steering, an inference-time framework for steering diffusion models with reward functions. FK steering works by sampling a system of multiple interacting diffusion processes, called particles, and resampling particles at intermediate steps based on scores computed using functions called potentials. Potentials are defined using rewards for intermediate states and are selected such that a high value indicates that the particle will yield a high-reward sample. We explore various choices of potentials, intermediate rewards, and samplers. We evaluate FK steering on text-to-image and text diffusion models. For steering text-to-image models with a human preference reward, we find that FK steering a 0.8B parameter model outperforms a 2.6B parameter fine-tuned model on prompt fidelity, with faster sampling and no training. For steering text diffusion models with rewards for text quality and specific text attributes, we find that FK steering generates lower perplexity, more linguistically acceptable outputs and enables gradient-free control of attributes like toxicity. Our results demonstrate that inference-time scaling and steering of diffusion models - even with off-the-shelf rewards - can provide significant sample quality gains and controllability benefits. Code is available at https://github.com/zacharyhorvitz/Fk-Diffusion-Steering .
♻ ☆ Beyond Scale: Small Language Models are Comparable to GPT-4 in Mental Health Understanding
The emergence of Small Language Models (SLMs) as privacy-preserving alternatives for sensitive applications raises a fundamental question about their inherent understanding capabilities compared to Large Language Models (LLMs). This paper investigates the mental health understanding capabilities of current SLMs through systematic evaluation across diverse classification tasks. Employing zero-shot and few-shot learning paradigms, we benchmark their performance against established LLM baselines to elucidate their relative strengths and limitations in this critical domain. We assess five state-of-the-art SLMs (Phi-3, Phi-3.5, Qwen2.5, Llama-3.2, Gemma2) against three LLMs (GPT-4, FLAN-T5-XXL, Alpaca-7B) on six mental health understanding tasks. Our findings reveal that SLMs achieve mean performance within 2\% of LLMs on binary classification tasks (F1 scores of 0.64 vs 0.66 in zero-shot settings), demonstrating notable competence despite orders of magnitude fewer parameters. Both model categories experience similar degradation on multi-class severity tasks (a drop of over 30\%), suggesting that nuanced clinical understanding challenges transcend model scale. Few-shot prompting provides substantial improvements for SLMs (up to 14.6\%), while LLM gains are more variable. Our work highlights the potential of SLMs in mental health understanding, showing they can be effective privacy-preserving tools for analyzing sensitive online text data. In particular, their ability to quickly adapt and specialize with minimal data through few-shot learning positions them as promising candidates for scalable mental health screening tools.
♻ ☆ Deep Binding of Language Model Virtual Personas: a Study on Approximating Political Partisan Misperceptions
Large language models (LLMs) are increasingly capable of simulating human behavior, offering cost-effective ways to estimate user responses to various surveys and polls. However, the questions in these surveys usually reflect socially understood attitudes: the patterns of attitudes of old/young, liberal/conservative, as understood by both members and non-members of those groups. It is not clear whether the LLM binding is \emph{deep}, meaning the LLM answers as a member of a particular in-group would, or \emph{shallow}, meaning the LLM responds as an out-group member believes an in-group member would. To explore this difference, we use questions that expose known in-group/out-group biases. This level of fidelity is critical for applying LLMs to various political science studies, including timely topics on polarization dynamics, inter-group conflict, and democratic backsliding. To this end, we propose a novel methodology for constructing virtual personas with synthetic user ``backstories" generated as extended, multi-turn interview transcripts. Our generated backstories are longer, rich in detail, and consistent in authentically describing a singular individual, compared to previous methods. We show that virtual personas conditioned on our backstories closely replicate human response distributions (up to an 87\% improvement as measured by Wasserstein Distance) and produce effect sizes that closely match those observed in the original studies of in-group/out-group biases. Altogether, our work extends the applicability of LLMs beyond estimating socially understood responses, enabling their use in a broader range of human studies.
comment: COLM 2025
♻ ☆ AIDE: Attribute-Guided MultI-Hop Data Expansion for Data Scarcity in Task-Specific Fine-tuning ACL 2025
Fine-tuning large language models (LLMs) for specific tasks requires diverse, high-quality training data. However, obtaining sufficient relevant data remains a significant challenge. Existing data synthesis methods either depend on extensive seed datasets or struggle to balance task relevance and data diversity. To address these challenges, we propose Attribute-guided multI-hop Data Expansion (AIDE), a novel data synthesis framework that uses a multi-hop process to expand very few seed data points while ensuring data diversity and task relevance. AIDE extracts the main topic and key knowledge attributes from the seeds to guide the synthesis steps. The process repeats for K hops, using the generated data as seeds. To prevent irrelevant data generation as the hop depth increases, AIDE incorporates a residual connection mechanism. Our empirical results show that AIDE enables fine-tuning of Mistral-7B, Llama-3.1-8B and Llama-3.2-3B from 10 seeds, surpassing the models fine-tuned on human curated data. Furthermore, AIDE outperforms state-of-the-art data synthesis methods, such as Evol-Instruct, by over 30% in task-specific fine-tuning. Code is available at https://github.com/Code4Graph/AIDE.
comment: Accepted for publication in ACL 2025. The official version will be available in the ACL Anthology
♻ ☆ Agentic Reasoning: A Streamlined Framework for Enhancing LLM Reasoning with Agentic Tools ACL 2025
We introduce Agentic Reasoning, a framework that enhances large language model (LLM) reasoning by integrating external tool-using agents. Agentic Reasoning dynamically leverages web search, code execution, and structured memory to address complex problems requiring deep research. A key innovation in our framework is the Mind-Map agent, which constructs a structured knowledge graph to store reasoning context and track logical relationships, ensuring coherence in long reasoning chains with extensive tool usage. Additionally, we conduct a comprehensive exploration of the Web-Search agent, leading to a highly effective search mechanism that surpasses all prior approaches. When deployed on DeepSeek-R1, our method achieves a new state-of-the-art (SOTA) among public models and delivers performance comparable to OpenAI Deep Research, the leading proprietary model in this domain. Extensive ablation studies validate the optimal selection of agentic tools and confirm the effectiveness of our Mind-Map and Web-Search agents in enhancing LLM reasoning. The code is at: https://github.com/theworldofagents/Agentic-Reasoning
comment: ACL 2025
♻ ☆ GDC Cohort Copilot: An AI Copilot for Curating Cohorts from the Genomic Data Commons
The Genomic Data Commons (GDC) provides access to high quality, harmonized cancer genomics data through a unified curation and analysis platform centered around patient cohorts. While GDC users can interactively create complex cohorts through the graphical Cohort Builder, users (especially new ones) may struggle to find specific cohort descriptors across hundreds of possible fields and properties. However, users may be better able to describe their desired cohort in free-text natural language. We introduce GDC Cohort Copilot, an open-source copilot tool for curating cohorts from the GDC. GDC Cohort Copilot automatically generates the GDC cohort filter corresponding to a user-input natural language description of their desired cohort, before exporting the cohort back to the GDC for further analysis. An interactive user interface allows users to further refine the generated cohort. We develop and evaluate multiple large language models (LLMs) for GDC Cohort Copilot and demonstrate that our locally-served, open-source GDC Cohort LLM achieves better results than GPT-4o prompting in generating GDC cohorts. We implement and share GDC Cohort Copilot as a containerized Gradio app on HuggingFace Spaces, available at https://huggingface.co/spaces/uc-ctds/GDC-Cohort-Copilot. GDC Cohort LLM weights are available at https://huggingface.co/uc-ctds. All source code is available at https://github.com/uc-cdis/gdc-cohort-copilot.
comment: 12 pages, 1 figure, 7 tables. v2 updated to reflect migration to HF Spaces
♻ ☆ DroidSpeak: KV Cache Sharing for Cross-LLM Communication and Multi-LLM Serving
Compound AI systems, such as agentic systems, are an emerging trend in large-scale enterprise settings, with multiple LLMs specialized for different users, tasks, and/or roles working together. In these scenarios, different models often process inputs that share the same context prefix. Although much work was done in the past to enable the reuse of prefix KV caches across inputs for a single model, how to enable one model to reuse the prefix KV caches of a different model remains an open question. We introduce DroidSpeak, the first distributed LLM inference system that enables KV cache reuse across distributed nodes running inference of different LLMs, so long as the LLMs have the same architecture. We present the first study that aims at understanding the impact of sharing KV caches across different LLMs, and if/when such sharing affects quality. Inspired by the findings, we present DroidSpeak, which selectively recomputes a few layers of the KV cache produced by another LLM and reuses the remaining layers, with negligible quality loss. Moreover, carefully pipelining the layer-wise re-computation and the loading of reused KV cache further improves the inference performance. Experiments on diverse datasets and model pairs demonstrate that DroidSpeak achieves up to 4x throughput improvement and about 3.1x faster prefill (time to first token), with negligible loss of quality in F1 scores, Rouge-L or code similarity score, compared to the baseline which does not allow any sharing across models.
Computer Vision and Pattern Recognition 15
Self-supervised Learning on Camera Trap Footage Yields a Strong Universal Face Embedder
Camera traps are revolutionising wildlife monitoring by capturing vast amounts of visual data; however, the manual identification of individual animals remains a significant bottleneck. This study introduces a fully self-supervised approach to learning robust chimpanzee face embeddings from unlabeled camera-trap footage. Leveraging the DINOv2 framework, we train Vision Transformers on automatically mined face crops, eliminating the need for identity labels. Our method demonstrates strong open-set re-identification performance, surpassing supervised baselines on challenging benchmarks such as Bossou, despite utilising no labelled data during training. This work underscores the potential of self-supervised learning in biodiversity monitoring and paves the way for scalable, non-invasive population studies.
comment: Accepted for publication. Project page, code and weights: https://www.robots.ox.ac.uk/~vgg/research/ChimpUFE/
☆ EmbRACE-3K: Embodied Reasoning and Action in Complex Environments
Recent advanced vision-language models(VLMs) have demonstrated strong performance on passive, offline image and video understanding tasks. However, their effectiveness in embodied settings, which require online interaction and active scene understanding remains limited. In such scenarios, an agent perceives the environment from a first-person perspective, with each action dynamically shaping subsequent observations. Even state-of-the-art models such as GPT-4o, Claude 3.5 Sonnet, and Gemini 2.5 Pro struggle in open-environment interactions, exhibiting clear limitations in spatial reasoning and long-horizon planning. To address this gap, we introduce EmRACE-3K, a dataset of over 3,000 language-guided tasks situated in diverse, photorealistic environments constructed using Unreal Engine and the UnrealCV-Zoo framework. The tasks encompass a wide range of embodied challenges, including navigation, object manipulation, and multi-stage goal execution. Each task unfolds as a multi-step trajectory, pairing first-person visual observations with high-level instructions, grounded actions, and natural language rationales that express the agent's intent at every step. Using EmRACE-3K, we establish a benchmark to evaluate the embodied reasoning capabilities of VLMs across three key dimensions: Exploration, Dynamic Spatial-Semantic Reasoning, and Multi-stage Goal Execution. In zero-shot settings, all models achieve success rates below 20%, underscoring the challenge posed by our benchmark and the current limitations of VLMs in interactive environments. To demonstrate the utility of EmRACE-3K, we further fine-tune Qwen2.5-VL-7B using supervised learning followed by reinforcement learning. This approach yields substantial improvements across all three challenge categories, highlighting the dataset's effectiveness in enabling the development of embodied reasoning capabilities.
comment: Project page: https://mxllc.github.io/EmbRACE-3K/
☆ Quantize-then-Rectify: Efficient VQ-VAE Training
Visual tokenizers are pivotal in multimodal large models, acting as bridges between continuous inputs and discrete tokens. Nevertheless, training high-compression-rate VQ-VAEs remains computationally demanding, often necessitating thousands of GPU hours. This work demonstrates that a pre-trained VAE can be efficiently transformed into a VQ-VAE by controlling quantization noise within the VAE's tolerance threshold. We present \textbf{Quantize-then-Rectify (ReVQ)}, a framework leveraging pre-trained VAEs to enable rapid VQ-VAE training with minimal computational overhead. By integrating \textbf{channel multi-group quantization} to enlarge codebook capacity and a \textbf{post rectifier} to mitigate quantization errors, ReVQ compresses ImageNet images into at most 512 tokens while sustaining competitive reconstruction quality (rFID = 1.06). Significantly, ReVQ reduces training costs by over two orders of magnitude relative to state-of-the-art approaches: ReVQ finishes full training on a single NVIDIA 4090 in approximately 22 hours, whereas comparable methods require 4.5 days on 32 A100 GPUs. Experimental results show that ReVQ achieves superior efficiency-reconstruction trade-offs.
☆ ScaffoldAvatar: High-Fidelity Gaussian Avatars with Patch Expressions SIGGRAPH 2025
Generating high-fidelity real-time animated sequences of photorealistic 3D head avatars is important for many graphics applications, including immersive telepresence and movies. This is a challenging problem particularly when rendering digital avatar close-ups for showing character's facial microfeatures and expressions. To capture the expressive, detailed nature of human heads, including skin furrowing and finer-scale facial movements, we propose to couple locally-defined facial expressions with 3D Gaussian splatting to enable creating ultra-high fidelity, expressive and photorealistic 3D head avatars. In contrast to previous works that operate on a global expression space, we condition our avatar's dynamics on patch-based local expression features and synthesize 3D Gaussians at a patch level. In particular, we leverage a patch-based geometric 3D face model to extract patch expressions and learn how to translate these into local dynamic skin appearance and motion by coupling the patches with anchor points of Scaffold-GS, a recent hierarchical scene representation. These anchors are then used to synthesize 3D Gaussians on-the-fly, conditioned by patch-expressions and viewing direction. We employ color-based densification and progressive training to obtain high-quality results and faster convergence for high resolution 3K training images. By leveraging patch-level expressions, ScaffoldAvatar consistently achieves state-of-the-art performance with visually natural motion, while encompassing diverse facial expressions and styles in real time.
comment: (SIGGRAPH 2025) Paper Video: https://youtu.be/VyWkgsGdbkk Project Page: https://shivangi-aneja.github.io/projects/scaffoldavatar/
☆ Scene-Aware Conversational ADAS with Generative AI for Real-Time Driver Assistance
While autonomous driving technologies continue to advance, current Advanced Driver Assistance Systems (ADAS) remain limited in their ability to interpret scene context or engage with drivers through natural language. These systems typically rely on predefined logic and lack support for dialogue-based interaction, making them inflexible in dynamic environments or when adapting to driver intent. This paper presents Scene-Aware Conversational ADAS (SC-ADAS), a modular framework that integrates Generative AI components including large language models, vision-to-text interpretation, and structured function calling to enable real-time, interpretable, and adaptive driver assistance. SC-ADAS supports multi-turn dialogue grounded in visual and sensor context, allowing natural language recommendations and driver-confirmed ADAS control. Implemented in the CARLA simulator with cloud-based Generative AI, the system executes confirmed user intents as structured ADAS commands without requiring model fine-tuning. We evaluate SC-ADAS across scene-aware, conversational, and revisited multi-turn interactions, highlighting trade-offs such as increased latency from vision-based context retrieval and token growth from accumulated dialogue history. These results demonstrate the feasibility of combining conversational reasoning, scene perception, and modular ADAS control to support the next generation of intelligent driver assistance.
☆ National level satellite-based crop field inventories in smallholder landscapes
The design of science-based policies to improve the sustainability of smallholder agriculture is challenged by a limited understanding of fundamental system properties, such as the spatial distribution of active cropland and field size. We integrate very high spatial resolution (1.5 m) Earth observation data and deep transfer learning to derive crop field delineations in complex agricultural systems at the national scale, while maintaining minimum reference data requirements and enhancing transferability. We provide the first national-level dataset of 21 million individual fields for Mozambique (covering ~800,000 km2) for 2023. Our maps separate active cropland from non-agricultural land use with an overall accuracy of 93% and balanced omission and commission errors. Field-level spatial agreement reached median intersection over union (IoU) scores of 0.81, advancing the state-of-the-art in large-area field delineation in complex smallholder systems. The active cropland maps capture fragmented rural regions with low cropland shares not yet identified in global land cover or cropland maps. These regions are mostly located in agricultural frontier regions which host 7-9% of the Mozambican population. Field size in Mozambique is very low overall, with half of the fields being smaller than 0.16 ha, and 83% smaller than 0.5 ha. Mean field size at aggregate spatial resolution (0.05{\deg}) is 0.32 ha, but it varies strongly across gradients of accessibility, population density, and net forest cover change. This variation reflects a diverse set of actors, ranging from semi-subsistence smallholder farms to medium-scale commercial farming, and large-scale farming operations. Our results highlight that field size is a key indicator relating to socio-economic and environmental outcomes of agriculture (e.g., food production, livelihoods, deforestation, biodiversity), as well as their trade-offs.
☆ Cameras as Relative Positional Encoding
Transformers are increasingly prevalent for multi-view computer vision tasks, where geometric relationships between viewpoints are critical for 3D perception. To leverage these relationships, multi-view transformers must use camera geometry to ground visual tokens in 3D space. In this work, we compare techniques for conditioning transformers on cameras: token-level raymap encodings, attention-level relative pose encodings, and a new relative encoding we propose -- Projective Positional Encoding (PRoPE) -- that captures complete camera frustums, both intrinsics and extrinsics, as a relative positional encoding. Our experiments begin by showing how relative camera conditioning improves performance in feedforward novel view synthesis, with further gains from PRoPE. This holds across settings: scenes with both shared and varying intrinsics, when combining token- and attention-level conditioning, and for generalization to inputs with out-of-distribution sequence lengths and camera intrinsics. We then verify that these benefits persist for different tasks, stereo depth estimation and discriminative spatial cognition, as well as larger model sizes.
comment: Project Page: https://www.liruilong.cn/prope/
☆ BenchReAD: A systematic benchmark for retinal anomaly detection MICCAI 2025
Retinal anomaly detection plays a pivotal role in screening ocular and systemic diseases. Despite its significance, progress in the field has been hindered by the absence of a comprehensive and publicly available benchmark, which is essential for the fair evaluation and advancement of methodologies. Due to this limitation, previous anomaly detection work related to retinal images has been constrained by (1) a limited and overly simplistic set of anomaly types, (2) test sets that are nearly saturated, and (3) a lack of generalization evaluation, resulting in less convincing experimental setups. Furthermore, existing benchmarks in medical anomaly detection predominantly focus on one-class supervised approaches (training only with negative samples), overlooking the vast amounts of labeled abnormal data and unlabeled data that are commonly available in clinical practice. To bridge these gaps, we introduce a benchmark for retinal anomaly detection, which is comprehensive and systematic in terms of data and algorithm. Through categorizing and benchmarking previous methods, we find that a fully supervised approach leveraging disentangled representations of abnormalities (DRA) achieves the best performance but suffers from significant drops in performance when encountering certain unseen anomalies. Inspired by the memory bank mechanisms in one-class supervised learning, we propose NFM-DRA, which integrates DRA with a Normal Feature Memory to mitigate the performance degradation, establishing a new SOTA. The benchmark is publicly available at https://github.com/DopamineLcy/BenchReAD.
comment: MICCAI 2025
☆ The Power of Certainty: How Confident Models Lead to Better Segmentation
Deep learning models have been proposed for automatic polyp detection and precise segmentation of polyps during colonoscopy procedures. Although these state-of-the-art models achieve high performance, they often require a large number of parameters. Their complexity can make them prone to overfitting, particularly when trained on biased datasets, and can result in poor generalization across diverse datasets. Knowledge distillation and self-distillation are proposed as promising strategies to mitigate the limitations of large, over-parameterized models. These approaches, however, are resource-intensive, often requiring multiple models and significant memory during training. We propose a confidence-based self-distillation approach that outperforms state-of-the-art models by utilizing only previous iteration data storage during training, without requiring extra computation or memory usage during testing. Our approach calculates the loss between the previous and current iterations within a batch using a dynamic confidence coefficient. To evaluate the effectiveness of our approach, we conduct comprehensive experiments on the task of polyp segmentation. Our approach outperforms state-of-the-art models and generalizes well across datasets collected from multiple clinical centers. The code will be released to the public once the paper is accepted.
comment: 9 pages, 3 figures
♻ ☆ Visual Test-time Scaling for GUI Agent Grounding ICCV2025
We introduce RegionFocus, a visual test-time scaling approach for Vision Language Model Agents. Understanding webpages is challenging due to the visual complexity of GUI images and the large number of interface elements, making accurate action selection difficult. Our approach dynamically zooms in on relevant regions, reducing background clutter and improving grounding accuracy. To support this process, we propose an image-as-map mechanism that visualizes key landmarks at each step, providing a transparent action record and enables the agent to effectively choose among action candidates. Even with a simple region selection strategy, we observe significant performance gains of 28+\% on Screenspot-pro and 24+\% on WebVoyager benchmarks on top of two state-of-the-art open vision language model agents, UI-TARS and Qwen2.5-VL, highlighting the effectiveness of visual test-time scaling in interactive settings. We achieve a new state-of-the-art grounding performance of 61.6\% on the ScreenSpot-Pro benchmark by applying RegionFocus to a Qwen2.5-VL-72B model. Our code will be released publicly at https://github.com/tiangeluo/RegionFocus.
comment: ICCV2025, https://github.com/tiangeluo/RegionFocus
♻ ☆ UniQA: Unified Vision-Language Pre-training for Image Quality and Aesthetic Assessment
Image Quality Assessment (IQA) and Image Aesthetic Assessment (IAA) aim to simulate human subjective perception of image visual quality and aesthetic appeal. Despite distinct learning objectives, they have underlying interconnectedness due to consistent human assessment perception. In this paper, we propose Unified vision-language pre-training of Quality and Aesthetics (UniQA}), to extract useful and common representations from two tasks, thereby benefiting them simultaneously. However, the lack of text in the IQA datasets and the textual noise in the IAA datasets pose severe challenges for multimodal pre-training. To address this, we (1) utilize multimodal large language models (MLLMs) to generate high-quality text descriptions; (2) use the generated text for IAA as metadata to purify noisy IAA data. To effectively adapt the pre-trained UniQA to downstream tasks, we further propose a lightweight adapter that utilizes versatile cues to fully exploit the extensive knowledge of the pre-trained model. UniQA demonstrates high competitiveness in various image assessment tasks, including classical IQA and IAA tasks, few-label IQA, and other downstream tasks, showing promise as a foundational assessment model. Codes are available at https://github.com/zht8506/UniQA.
♻ ☆ Gamma: Toward Generic Image Assessment with Mixture of Assessment Experts
Image assessment aims to evaluate the quality and aesthetics of images and has been applied across various scenarios, such as natural and AIGC scenes. Existing methods mostly address these sub-tasks or scenes individually. While some works attempt to develop unified image assessment models, they have struggled to achieve satisfactory performance or cover a broad spectrum of assessment scenarios. In this paper, we present \textbf{Gamma}, a \textbf{G}eneric im\textbf{A}ge assess\textbf{M}ent model using \textbf{M}ixture of \textbf{A}ssessment Experts, which can effectively assess images from diverse scenes through mixed-dataset training. Achieving unified training in image assessment presents significant challenges due to annotation biases across different datasets. To address this issue, we first propose a Mixture of Assessment Experts (MoAE) module, which employs shared and adaptive experts to dynamically learn common and specific knowledge for different datasets, respectively. In addition, we introduce a Scene-based Differential Prompt (SDP) strategy, which uses scene-specific prompts to provide prior knowledge and guidance during the learning process, further boosting adaptation for various scenes. Our Gamma model is trained and evaluated on 12 datasets spanning 6 image assessment scenarios. Extensive experiments show that our unified Gamma outperforms other state-of-the-art mixed-training methods by significant margins while covering more scenes. Codes are available at https://github.com/zht8506/Gamma.
comment: Accepted to ACMMM 2025
♻ ☆ GaussianOcc: Fully Self-supervised and Efficient 3D Occupancy Estimation with Gaussian Splatting
We introduce GaussianOcc, a systematic method that investigates the two usages of Gaussian splatting for fully self-supervised and efficient 3D occupancy estimation in surround views. First, traditional methods for self-supervised 3D occupancy estimation still require ground truth 6D poses from sensors during training. To address this limitation, we propose Gaussian Splatting for Projection (GSP) module to provide accurate scale information for fully self-supervised training from adjacent view projection. Additionally, existing methods rely on volume rendering for final 3D voxel representation learning using 2D signals (depth maps, semantic maps), which is both time-consuming and less effective. We propose Gaussian Splatting from Voxel space (GSV) to leverage the fast rendering properties of Gaussian splatting. As a result, the proposed GaussianOcc method enables fully self-supervised (no ground truth pose) 3D occupancy estimation in competitive performance with low computational cost (2.7 times faster in training and 5 times faster in rendering). The relevant code is available in https://github.com/GANWANSHUI/GaussianOcc.git.
comment: Project page: https://ganwanshui.github.io/GaussianOcc/
♻ ☆ Alignment and Adversarial Robustness: Are More Human-Like Models More Secure? SP
A small but growing body of work has shown that machine learning models which better align with human vision have also exhibited higher robustness to adversarial examples, raising the question: can human-like perception make models more secure? If true generally, such mechanisms would offer new avenues toward robustness. In this work, we conduct a large-scale empirical analysis to systematically investigate the relationship between representational alignment and adversarial robustness. We evaluate 114 models spanning diverse architectures and training paradigms, measuring their neural and behavioral alignment and engineering task performance across 105 benchmarks as well as their adversarial robustness via AutoAttack. Our findings reveal that while average alignment and robustness exhibit a weak overall correlation, specific alignment benchmarks serve as strong predictors of adversarial robustness, particularly those that measure selectivity toward texture or shape. These results suggest that different forms of alignment play distinct roles in model robustness, motivating further investigation into how alignment-driven approaches can be leveraged to build more secure and perceptually-grounded vision models.
comment: Accepted to International Workshop on Security and Privacy-Preserving AI/ML (SPAIML) 2025
♻ ☆ On the Robustness Tradeoff in Fine-Tuning ICCV 2025
Fine-tuning has become the standard practice for adapting pre-trained models to downstream tasks. However, the impact on model robustness is not well understood. In this work, we characterize the robustness-accuracy trade-off in fine-tuning. We evaluate the robustness and accuracy of fine-tuned models over 6 benchmark datasets and 7 different fine-tuning strategies. We observe a consistent trade-off between adversarial robustness and accuracy. Peripheral updates such as BitFit are more effective for simple tasks -- over 75% above the average measured by the area under the Pareto frontiers on CIFAR-10 and CIFAR-100. In contrast, fine-tuning information-heavy layers, such as attention layers via Compacter, achieves a better Pareto frontier on more complex tasks -- 57.5% and 34.6% above the average on Caltech-256 and CUB-200, respectively. Lastly, we observe that the robustness of fine-tuning against out-of-distribution data closely tracks accuracy. These insights emphasize the need for robustness-aware fine-tuning to ensure reliable real-world deployments.
comment: Accepted to International Conference on Computer Vision, ICCV 2025
Information Retrieval 21
☆ Overcoming catastrophic forgetting in neural networks
Catastrophic forgetting is the primary challenge that hinders continual learning, which refers to a neural network ability to sequentially learn multiple tasks while retaining previously acquired knowledge. Elastic Weight Consolidation, a regularization-based approach inspired by synaptic consolidation in biological neural systems, has been used to overcome this problem. In this study prior research is replicated and extended by evaluating EWC in supervised learning settings using the PermutedMNIST and RotatedMNIST benchmarks. Through systematic comparisons with L2 regularization and stochastic gradient descent (SGD) without regularization, we analyze how different approaches balance knowledge retention and adaptability. Our results confirm what was shown in previous research, showing that EWC significantly reduces forgetting compared to naive training while slightly compromising learning efficiency on new tasks. Moreover, we investigate the impact of dropout regularization and varying hyperparameters, offering insights into the generalization of EWC across diverse learning scenarios. These results underscore EWC's potential as a viable solution for lifelong learning in neural networks.
comment: 7 pages, 5 figures, EE-411 Fundamentals of inference and learning course project
☆ SentiDrop: A Multi Modal Machine Learning model for Predicting Dropout in Distance Learning
School dropout is a serious problem in distance learning, where early detection is crucial for effective intervention and student perseverance. Predicting student dropout using available educational data is a widely researched topic in learning analytics. Our partner's distance learning platform highlights the importance of integrating diverse data sources, including socio-demographic data, behavioral data, and sentiment analysis, to accurately predict dropout risks. In this paper, we introduce a novel model that combines sentiment analysis of student comments using the Bidirectional Encoder Representations from Transformers (BERT) model with socio-demographic and behavioral data analyzed through Extreme Gradient Boosting (XGBoost). We fine-tuned BERT on student comments to capture nuanced sentiments, which were then merged with key features selected using feature importance techniques in XGBoost. Our model was tested on unseen data from the next academic year, achieving an accuracy of 84\%, compared to 82\% for the baseline model. Additionally, the model demonstrated superior performance in other metrics, such as precision and F1-score. The proposed method could be a vital tool in developing personalized strategies to reduce dropout rates and encourage student perseverance
comment: International Conference on Education and New Learning Technologies (2025)
☆ Am I on the Right Track? What Can Predicted Query Performance Tell Us about the Search Behaviour of Agentic RAG
Agentic Retrieval-Augmented Generation (RAG) is a new paradigm where the reasoning model decides when to invoke a retriever (as a "tool") when answering a question. This paradigm, exemplified by recent research works such as Search-R1, enables the model to decide when to search and obtain external information. However, the queries generated by such Agentic RAG models and the role of the retriever in obtaining high-quality answers remain understudied. To this end, this initial study examines the applicability of query performance prediction (QPP) within the recent Agentic RAG models Search-R1 and R1-Searcher. We find that applying effective retrievers can achieve higher answer quality within a shorter reasoning process. Moreover, the QPP estimates of the generated queries, used as an approximation of their retrieval quality, are positively correlated with the quality of the final answer. Ultimately, our work is a step towards adaptive retrieval within Agentic RAG, where QPP is used to inform the model if the retrieved results are likely to be useful.
☆ Text-to-Remote-Sensing-Image Retrieval beyond RGB Sources
Retrieving relevant imagery from vast satellite archives is crucial for applications like disaster response and long-term climate monitoring. However, most text-to-image retrieval systems are limited to RGB data, failing to exploit the unique physical information captured by other sensors, such as the all-weather structural sensitivity of Synthetic Aperture Radar (SAR) or the spectral signatures in optical multispectral data. To bridge this gap, we introduce CrisisLandMark, a new large-scale corpus of over 647,000 Sentinel-1 SAR and Sentinel-2 multispectral images paired with structured textual annotations for land cover, land use, and crisis events harmonized from authoritative land cover systems (CORINE and Dynamic World) and crisis-specific sources. We then present CLOSP (Contrastive Language Optical SAR Pretraining), a novel framework that uses text as a bridge to align unpaired optical and SAR images into a unified embedding space. Our experiments show that CLOSP achieves a new state-of-the-art, improving retrieval nDGC by 54% over existing models. Additionally, we find that the unified training strategy overcomes the inherent difficulty of interpreting SAR imagery by transferring rich semantic knowledge from the optical domain with indirect interaction. Furthermore, GeoCLOSP, which integrates geographic coordinates into our framework, creates a powerful trade-off between generality and specificity: while the CLOSP excels at general semantic tasks, the GeoCLOSP becomes a specialized expert for retrieving location-dependent crisis events and rare geographic features. This work highlights that the integration of diverse sensor data and geographic context is essential for unlocking the full potential of remote sensing archives.
☆ Riding the Carousel: The First Extensive Eye Tracking Analysis of Browsing Behavior in Carousel Recommenders
Carousels have become the de-facto interface in online services. However, there is a lack of research in carousels, particularly examining how recommender systems may be designed differently than the traditional single-list interfaces. One of the key elements for understanding how to design a system for a particular interface is understanding how users browse. For carousels, users may browse in a number of different ways due to the added complexity of multiple topic defined-lists and swiping to see more items. Eye tracking is the key to understanding user behavior by providing valuable, direct information on how users see and navigate. In this work, we provide the first extensive analysis of the eye tracking behavior in carousel recommenders under the free-browsing setting. To understand how users browse, we examine the following research questions : 1) where do users start browsing, 2) how do users transition from item to item within the same carousel and across carousels, and 3) how does genre preference impact transitions? This work addresses a gap in the field and provides the first extensive empirical results of eye tracked browsing behavior in carousels for improving recommenders. Taking into account the insights learned from the above questions, our final contribution is to provide suggestions to help carousel recommender system designers optimize their systems for user browsing behavior. The most important suggestion being to reorder the ranked item positions to account for browsing after swiping.These contributions aim not only to help improve current systems, but also to encourage and allow the design of new user models, systems, and metrics that are better suited to the complexity of carousel interfaces.
☆ User Long-Term Multi-Interest Retrieval Model for Recommendation
User behavior sequence modeling, which captures user interest from rich historical interactions, is pivotal for industrial recommendation systems. Despite breakthroughs in ranking-stage models capable of leveraging ultra-long behavior sequences with length scaling up to thousands, existing retrieval models remain constrained to sequences of hundreds of behaviors due to two main challenges. One is strict latency budget imposed by real-time service over large-scale candidate pool. The other is the absence of target-aware mechanisms and cross-interaction architectures, which prevent utilizing ranking-like techniques to simplify long sequence modeling. To address these limitations, we propose a new framework named User Long-term Multi-Interest Retrieval Model(ULIM), which enables thousand-scale behavior modeling in retrieval stages. ULIM includes two novel components: 1)Category-Aware Hierarchical Dual-Interest Learning partitions long behavior sequences into multiple category-aware subsequences representing multi-interest and jointly optimizes long-term and short-term interests within specific interest cluster. 2)Pointer-Enhanced Cascaded Category-to-Item Retrieval introduces Pointer-Generator Interest Network(PGIN) for next-category prediction, followed by next-item retrieval upon the top-K predicted categories. Comprehensive experiments on Taobao dataset show that ULIM achieves substantial improvement over state-of-the-art methods, and brings 5.54% clicks, 11.01% orders and 4.03% GMV lift for Taobaomiaosha, a notable mini-app of Taobao.
☆ PRISM: Fine-Grained Paper-to-Paper Retrieval with Multi-Aspect-Aware Query Optimization
Scientific paper retrieval, particularly framed as document-to-document retrieval, aims to identify relevant papers in response to a long-form query paper, rather than a short query string. Previous approaches to this task have focused on abstracts, embedding them into dense vectors as surrogates for full documents and calculating similarity across them, although abstracts provide only sparse and high-level summaries. To address this, we propose PRISM, a novel document-to-document retrieval method that introduces multiple, fine-grained representations for both the query and candidate papers. In particular, each query paper is decomposed into multiple aspect-specific views and individually embedded, which are then matched against candidate papers similarity segmented to consider their multifaceted dimensions. Moreover, we present SciFullBench, a novel benchmark in which the complete and segmented context of full papers for both queries and candidates is available. Then, experimental results show that PRISM improves performance by an average of 4.3% over existing retrieval baselines.
☆ SLIF-MR: Self-loop Iterative Fusion of Heterogeneous Auxiliary Information for Multimodal Recommendation
Knowledge graphs (KGs) and multimodal item information, which respectively capture relational and attribute features, play a crucial role in improving recommender system accuracy. Recent studies have attempted to integrate them via multimodal knowledge graphs (MKGs) to further enhance recommendation performance. However, existing methods typically freeze the MKG structure during training, which limits the full integration of structural information from heterogeneous graphs (e.g., KG and user-item interaction graph), and results in sub-optimal performance. To address this challenge, we propose a novel framework, termed Self-loop Iterative Fusion of Heterogeneous Auxiliary Information for Multimodal Recommendation (SLIF-MR), which leverages item representations from previous training epoch as feedback signals to dynamically optimize the heterogeneous graph structures composed of KG, multimodal item feature graph, and user-item interaction graph. Through this iterative fusion mechanism, both user and item representations are refined, thus improving the final recommendation performance. Specifically, based on the feedback item representations, SLIF-MR constructs an item-item correlation graph, then integrated into the establishment process of heterogeneous graphs as additional new structural information in a self-loop manner. Consequently, the internal structures of heterogeneous graphs are updated with the feedback item representations during training. Moreover, a semantic consistency learning strategy is proposed to align heterogeneous item representations across modalities. The experimental results show that SLIF-MR significantly outperforms existing methods, particularly in terms of accuracy and robustness.
comment: 10 pages,7 figures
☆ Non-parametric Graph Convolution for Re-ranking in Recommendation Systems RecSys2025
Graph knowledge has been proven effective in enhancing item rankings in recommender systems (RecSys), particularly during the retrieval stage. However, its application in the ranking stage, especially when richer contextual information in user-item interactions is available, remains underexplored. A major challenge lies in the substantial computational cost associated with repeatedly retrieving neighborhood information from billions of items stored in distributed systems. This resource-intensive requirement makes it difficult to scale graph-based methods in practical RecSys. To bridge this gap, we first demonstrate that incorporating graphs in the ranking stage improves ranking qualities. Notably, while the improvement is evident, we show that the substantial computational overheads entailed by graphs are prohibitively expensive for real-world recommendations. In light of this, we propose a non-parametric strategy that utilizes graph convolution for re-ranking only during test time. Our strategy circumvents the notorious computational overheads from graph convolution during training, and utilizes structural knowledge hidden in graphs on-the-fly during testing. It can be used as a plug-and-play module and easily employed to enhance the ranking ability of various ranking layers of a real-world RecSys with significantly reduced computational overhead. Through comprehensive experiments across four benchmark datasets with varying levels of sparsity, we demonstrate that our strategy yields noticeable improvements (i.e., 8.1% on average) during testing time with little to no additional computational overheads (i.e., 0.5 on average). Code: https://github.com/zyouyang/RecSys2025_NonParamGC.git
comment: Accepted to RecSys2025 Main
☆ MixLoRA-DSI: Dynamically Expandable Mixture-of-LoRA Experts for Rehearsal-Free Generative Retrieval over Dynamic Corpora
Continually updating model-based indexes in generative retrieval with new documents remains challenging, as full retraining is computationally expensive and impractical under resource constraints. We propose MixLoRA-DSI, a novel framework that combines an expandable mixture of Low-Rank Adaptation experts with a layer-wise out-of-distribution (OOD)-driven expansion strategy. Instead of allocating new experts for each new corpus, our proposed expansion strategy enables sublinear parameter growth by selectively introducing new experts only when significant number of OOD documents are detected. Experiments on NQ320k and MS MARCO Passage demonstrate that MixLoRA-DSI outperforms full-model update baselines, with minimal parameter overhead and substantially lower training costs.
☆ Automated Thematic Analyses Using LLMs: Xylazine Wound Management Social Media Chatter Use Case
Background Large language models (LLMs) face challenges in inductive thematic analysis, a task requiring deep interpretive and domain-specific expertise. We evaluated the feasibility of using LLMs to replicate expert-driven thematic analysis of social media data. Methods Using two temporally non-intersecting Reddit datasets on xylazine (n=286 and n=686, for model optimization and validation, respectively) with twelve expert-derived themes, we evaluated five LLMs against expert coding. We modeled the task as a series of binary classifications, rather than a single, multi-label classification, employing zero-, single-, and few-shot prompting strategies and measuring performance via accuracy, precision, recall, and F1-score. Results On the validation set, GPT-4o with two-shot prompting performed best (accuracy: 90.9%; F1-score: 0.71). For high-prevalence themes, model-derived thematic distributions closely mirrored expert classifications (e.g., xylazine use: 13.6% vs. 17.8%; MOUD use: 16.5% vs. 17.8%). Conclusions Our findings suggest that few-shot LLM-based approaches can automate thematic analyses, offering a scalable supplement for qualitative research. Keywords: thematic analysis, large language models, natural language processing, qualitative analysis, social media, prompt engineering, public health
comment: Pages: 19, Abstract word count: 151 words, Manuscript word count: 2185 words, References: 14, Figures: 3, Tables: 2
☆ Applying Text Embedding Models for Efficient Analysis in Labeled Property Graphs
Labeled property graphs often contain rich textual attributes that can enhance analytical tasks when properly leveraged. This work explores the use of pretrained text embedding models to enable efficient semantic analysis in such graphs. By embedding textual node and edge properties, we support downstream tasks including node classification and relation prediction with improved contextual understanding. Our approach integrates language model embeddings into the graph pipeline without altering its structure, demonstrating that textual semantics can significantly enhance the accuracy and interpretability of property graph analysis.
☆ Access Control for Information-Theoretically Secure Key-Document Stores VLDB 2025
This paper presents a novel key-based access control technique for secure outsourcing key-value stores where values correspond to documents that are indexed and accessed using keys. The proposed approach adopts Shamir's secret-sharing that offers unconditional or information-theoretic security. It supports keyword-based document retrieval while preventing leakage of the data, access rights of users, or the size (\textit{i}.\textit{e}., volume of the output that satisfies a query). The proposed approach allows servers to detect (and abort) malicious clients from gaining unauthorized access to data, and prevents malicious servers from altering data undetected while ensuring efficient access -- it takes 231.5ms over 5,000 keywords across 500,000 files.
comment: An extended abstract of this version has been accepted in VLDB 2025
☆ Extracting Document Relations from Search Corpus by Marginalizing over User Queries
Understanding relationships between documents in large-scale corpora is essential for knowledge discovery and information organization. However, existing approaches rely heavily on manual annotation or predefined relationship taxonomies. We propose EDR-MQ (Extracting Document Relations by Marginalizing over User Queries), a novel framework that discovers document relationships through query marginalization. EDR-MQ is based on the insight that strongly related documents often co-occur in results across diverse user queries, enabling us to estimate joint probabilities between document pairs by marginalizing over a collection of queries. To enable this query marginalization approach, we develop Multiply Conditioned Retrieval-Augmented Generation (MC-RAG), which employs conditional retrieval where subsequent document retrievals depend on previously retrieved content. By observing co-occurrence patterns across diverse queries, EDR-MQ estimates joint probabilities between document pairs without requiring labeled training data or predefined taxonomies. Experimental results show that our query marginalization approach successfully identifies meaningful document relationships, revealing topical clusters, evidence chains, and cross-domain connections that are not apparent through traditional similarity-based methods. Our query-driven framework offers a practical approach to document organization that adapts to different user perspectives and information needs.
comment: 9 pages, 6 figures
♻ ☆ Beyond classical and contemporary models: a transformative AI framework for student dropout prediction in distance learning using RAG, Prompt engineering, and Cross-modal fusion
Student dropout in distance learning remains a critical challenge, with profound societal and economic consequences. While classical machine learning models leverage structured socio-demographic and behavioral data, they often fail to capture the nuanced emotional and contextual factors embedded in unstructured student interactions. This paper introduces a transformative AI framework that redefines dropout prediction through three synergistic innovations: Retrieval-Augmented Generation (RAG) for domain-specific sentiment analysis, prompt engineering to decode academic stressors,and cross-modal attention fusion to dynamically align textual, behavioral, and socio-demographic insights. By grounding sentiment analysis in a curated knowledge base of pedagogical content, our RAG-enhanced BERT model interprets student comments with unprecedented contextual relevance, while optimized prompts isolate indicators of academic distress (e.g., "isolation," "workload anxiety"). A cross-modal attention layer then fuses these insights with temporal engagement patterns, creating holistic risk pro-files. Evaluated on a longitudinal dataset of 4 423 students, the framework achieves 89% accuracy and an F1-score of 0.88, outperforming conventional models by 7% and reducing false negatives by 21%. Beyond prediction, the system generates interpretable interventions by retrieving contextually aligned strategies (e.g., mentorship programs for isolated learners). This work bridges the gap between predictive analytics and actionable pedagogy, offering a scalable solution to mitigate dropout risks in global education systems
comment: 13 pages, 8 figures, 1 Algorithms, 17th International Conference on Education and New Learning Technologies,: 30 June-2 July, 2025 Location: Palma, Spain
♻ ☆ Modeling Item-Level Dynamic Variability with Residual Diffusion for Bundle Recommendation
Existing solutions for bundle recommendation(BR) have achieved remarkable effectiveness for predicting the user's preference for prebuilt bundles. However, bundle-item(B-I) affiliation will vary dynamically in real scenarios. For example, a bundle themed as 'casual outfit', may add 'hat' or remove 'watch' due to factors such as seasonal variations, changes in user pes or inventory adjustments. Our empirical study demonstrates that the performance of mainstream BR models will fluctuate or even decline regarding item-level variability. This paper makes the first attempt to referencaddress the above problem and proposes a novel Residual Diffusion for Bundle Recommendation(RDiffBR) as a model-agnostic generative framework which can assist a BR model in adapting this scenario. During the initial training of the BR model, RDiffBR employs a residual diffusion model to process the item-level bundle embeddings which are generated by BR model to represent bundle theme via a forward-reverse process. In the inference stage, RDiffBR reverses item-level bundle embeddings obtained by the well-trained bundle model under B-I variability scenarios to generate the effective item-level bundle embeddings. In particular, the residual connection in our residual approximator significantly enhances item-level bundle embeddings generation ability of BR models. Experiments on six BR models and four public datasets from different domains show that RDiffBR improves the performance of Recall and NDCG of backbone BR models by up to 23%, while only increased training time about 4%.Codes and datasets are available at https://anonymous.4open.science/r/RDiffBR.
♻ ☆ Interpret the Internal States of Recommendation Model with Sparse Autoencoder
Recommendation model interpretation aims to reveal models' calculation process, enhancing their transparency, interpretability, and trustworthiness by clarifying the relationships between inputs, model activations, and outputs. However, the complex, often opaque nature of deep learning models complicates interpretation, and most existing methods are tailored to specific model architectures, limiting their generalizability across different types of recommendation models. To address these challenges, we propose RecSAE, an automated and generalizable probing framework that interprets Recommenders with Sparse AutoEncoder. It extracts interpretable latents from the internal states of recommendation models and links them to semantic concepts for interpretation. RecSAE does not alter original models during interpretation and also enables targeted de-biasing to models based on interpreted results. Specifically, RecSAE operates in three steps: First, it probes activations before the prediction layer to capture internal representations. Next, the RecSAE module is trained on these activations with a larger latent space and sparsity constraints, making the RecSAE latents more mono-semantic than the original model activations. Thirdly, RecSAE utilizes a language model to construct concept descriptions with confidence scores based on the relationships between latent activations and recommendation outputs. Experiments on three types of models (general, graph-based, and sequential) with three widely used datasets demonstrate the effectiveness and generalization of RecSAE framework. The interpreted concepts are further validated by human experts, showing strong alignment with human perception. Overall, RecSAE serves as a novel step in both model-level interpretations to various types of recommenders without affecting their functions and offering the potential for targeted tuning of models.
♻ ☆ Graph and Sequential Neural Networks in Session-based Recommendation: A Survey
Recent years have witnessed the remarkable success of recommendation systems (RSs) in alleviating the information overload problem. As a new paradigm of RSs, session-based recommendation (SR) specializes in users' short-term preference capture and aims to provide a more dynamic and timely recommendation based on the ongoing interacted actions. In this survey, we will give a comprehensive overview of the recent works on SR. First, we clarify the definitions of various SR tasks and introduce the characteristics of session-based recommendation against other recommendation tasks. Then, we summarize the existing methods in two categories: sequential neural network based methods and graph neural network (GNN) based methods. The standard frameworks and technical are also introduced. Finally, we discuss the challenges of SR and new research directions in this area.
♻ ☆ GR-LLMs: Recent Advances in Generative Recommendation Based on Large Language Models
In the past year, Generative Recommendations (GRs) have undergone substantial advancements, especially in leveraging the powerful sequence modeling and reasoning capabilities of Large Language Models (LLMs) to enhance overall recommendation performance. LLM-based GRs are forming a new paradigm that is distinctly different from discriminative recommendations, showing strong potential to replace traditional recommendation systems heavily dependent on complex hand-crafted features. In this paper, we provide a comprehensive survey aimed at facilitating further research of LLM-based GRs. Initially, we outline the general preliminaries and application cases of LLM-based GRs. Subsequently, we introduce the main considerations when LLM-based GRs are applied in real industrial scenarios. Finally, we explore promising directions for LLM-based GRs. We hope that this survey contributes to the ongoing advancement of the GR domain.
comment: 8 pages, 3 figures
♻ ☆ Towards Efficient Quantity Retrieval from Text:An Approach via Description Parsing and Weak Supervision
Quantitative facts are continually generated by companies and governments, supporting data-driven decision-making. While common facts are structured, many long-tail quantitative facts remain buried in unstructured documents, making them difficult to access. We propose the task of Quantity Retrieval: given a description of a quantitative fact, the system returns the relevant value and supporting evidence. Understanding quantity semantics in context is essential for this task. We introduce a framework based on description parsing that converts text into structured (description, quantity) pairs for effective retrieval. To improve learning, we construct a large paraphrase dataset using weak supervision based on quantity co-occurrence. We evaluate our approach on a large corpus of financial annual reports and a newly annotated quantity description dataset. Our method significantly improves top-1 retrieval accuracy from 30.98 percent to 64.66 percent.
comment: Extended version of the paper accepted in DEXA 2025
♻ ☆ External Large Foundation Model: How to Efficiently Serve Trillions of Parameters for Online Ads Recommendation WWW
Ads recommendation is a prominent service of online advertising systems and has been actively studied. Recent studies indicate that scaling-up and advanced design of the recommendation model can bring significant performance improvement. However, with a larger model scale, such prior studies have a significantly increasing gap from industry as they often neglect two fundamental challenges in industrial-scale applications. First, training and inference budgets are restricted for the model to be served, exceeding which may incur latency and impair user experience. Second, large-volume data arrive in a streaming mode with data distributions dynamically shifting, as new users/ads join and existing users/ads leave the system. We propose the External Large Foundation Model (ExFM) framework to address the overlooked challenges. Specifically, we develop external distillation and a data augmentation system (DAS) to control the computational cost of training/inference while maintaining high performance. We design the teacher in a way like a foundation model (FM) that can serve multiple students as vertical models (VMs) to amortize its building cost. We propose Auxiliary Head and Student Adapter to mitigate the data distribution gap between FM and VMs caused by the streaming data issue. Comprehensive experiments on internal industrial-scale applications and public datasets demonstrate significant performance gain by ExFM.
comment: Accepted by the ACM Web Conference (WWW) 2025 Industrial Track as Oral Presentation
Machine Learning 150
Self-supervised Learning on Camera Trap Footage Yields a Strong Universal Face Embedder
Camera traps are revolutionising wildlife monitoring by capturing vast amounts of visual data; however, the manual identification of individual animals remains a significant bottleneck. This study introduces a fully self-supervised approach to learning robust chimpanzee face embeddings from unlabeled camera-trap footage. Leveraging the DINOv2 framework, we train Vision Transformers on automatically mined face crops, eliminating the need for identity labels. Our method demonstrates strong open-set re-identification performance, surpassing supervised baselines on challenging benchmarks such as Bossou, despite utilising no labelled data during training. This work underscores the potential of self-supervised learning in biodiversity monitoring and paves the way for scalable, non-invasive population studies.
comment: Accepted for publication. Project page, code and weights: https://www.robots.ox.ac.uk/~vgg/research/ChimpUFE/
☆ Quantize-then-Rectify: Efficient VQ-VAE Training
Visual tokenizers are pivotal in multimodal large models, acting as bridges between continuous inputs and discrete tokens. Nevertheless, training high-compression-rate VQ-VAEs remains computationally demanding, often necessitating thousands of GPU hours. This work demonstrates that a pre-trained VAE can be efficiently transformed into a VQ-VAE by controlling quantization noise within the VAE's tolerance threshold. We present \textbf{Quantize-then-Rectify (ReVQ)}, a framework leveraging pre-trained VAEs to enable rapid VQ-VAE training with minimal computational overhead. By integrating \textbf{channel multi-group quantization} to enlarge codebook capacity and a \textbf{post rectifier} to mitigate quantization errors, ReVQ compresses ImageNet images into at most 512 tokens while sustaining competitive reconstruction quality (rFID = 1.06). Significantly, ReVQ reduces training costs by over two orders of magnitude relative to state-of-the-art approaches: ReVQ finishes full training on a single NVIDIA 4090 in approximately 22 hours, whereas comparable methods require 4.5 days on 32 A100 GPUs. Experimental results show that ReVQ achieves superior efficiency-reconstruction trade-offs.
☆ Disentangling Neural Disjunctive Normal Form Models
Neural Disjunctive Normal Form (DNF) based models are powerful and interpretable approaches to neuro-symbolic learning and have shown promising results in classification and reinforcement learning settings without prior knowledge of the tasks. However, their performance is degraded by the thresholding of the post-training symbolic translation process. We show here that part of the performance degradation during translation is due to its failure to disentangle the learned knowledge represented in the form of the networks' weights. We address this issue by proposing a new disentanglement method; by splitting nodes that encode nested rules into smaller independent nodes, we are able to better preserve the models' performance. Through experiments on binary, multiclass, and multilabel classification tasks (including those requiring predicate invention), we demonstrate that our disentanglement method provides compact and interpretable logical representations for the neural DNF-based models, with performance closer to that of their pre-translation counterparts. Our code is available at https://github.com/kittykg/disentangling-ndnf-classification.
comment: Accepted at NeSy 2025
☆ Fusing LLM Capabilities with Routing Data
The rapid advancement of large language models (LLMs) has created a vibrant ecosystem of diverse architectures, each with unique strengths due to differences in design, training data, and objectives. However, most applications still rely on a single backend model, limiting coverage of capabilities and leading to inefficiencies in performance and token cost when tackling complex tasks. We highlight an underexploited opportunity: LLM routing data, produced when hosting platforms route diverse queries to different models, which can reveal comparative strengths across tasks. To address this, we propose FusionBench, a comprehensive routing benchmark covering 14 tasks across five domains with 20 open-source LLMs (8B to 671B parameters), capturing 103M tokens and summarizing reusable thought templates from top models. Building on this, we introduce FusionFactory, a systematic fusion framework with three levels: (1) query-level fusion, tailoring routers for each query using both direct responses and reasoning-augmented outputs; (2) thought-level fusion, leveraging abstract templates derived from top-performing LLMs' answers to similar queries; and (3) model-level fusion, transferring capabilities between models via distillation, using top responses or highest judge scores as training data. Experiments show FusionFactory consistently outperforms the best individual LLM across all 14 benchmarks, with optimal fusion configurations varying by benchmark, demonstrating the value of systematic LLM fusion in harnessing complementary strengths and improving overall performance.
☆ Graph World Model
World models (WMs) demonstrate strong capabilities in prediction, generation, and planning tasks. Existing WMs primarily focus on unstructured data and cannot leverage the ubiquitous structured data, often represented as graphs, in the digital world. While multiple graph foundation models have been proposed, they focus on graph learning tasks and cannot extend to diverse multi-modal data and interdisciplinary tasks. To address these challenges, we propose the Graph World Model (GWM), a world model that supports both unstructured and graph-structured states with multi-modal information and represents diverse tasks as actions. The core of a GWM is a generic message-passing algorithm to aggregate structured information, either over a unified multi-modal token space by converting multi-modal data into text (GWM-T) or a unified multi-modal embedding space by modality-specific encoders (GWM-E). Notably, GWM introduces action nodes to support diverse tasks, where action nodes are linked to other nodes via direct reference or similarity computation. Extensive experiments on six tasks from diverse domains, including multi-modal generation and matching, recommendation, graph prediction, multi-agent, retrieval-augmented generation, and planning and optimization, show that the same GWM outperforms or matches domain-specific baselines' performance, benefits from multi-hop structures, and demonstrates strong zero-shot/few-shot capabilities on unseen new tasks. Our code for GWM is released at https://github.com/ulab-uiuc/GWM.
☆ On the Performance of Differentially Private Optimization with Heavy-Tail Class Imbalance
In this work, we analyze the optimization behaviour of common private learning optimization algorithms under heavy-tail class imbalanced distribution. We show that, in a stylized model, optimizing with Gradient Descent with differential privacy (DP-GD) suffers when learning low-frequency classes, whereas optimization algorithms that estimate second-order information do not. In particular, DP-AdamBC that removes the DP bias from estimating loss curvature is a crucial component to avoid the ill-condition caused by heavy-tail class imbalance, and empirically fits the data better with $\approx8\%$ and $\approx5\%$ increase in training accuracy when learning the least frequent classes on both controlled experiments and real data respectively.
☆ Reasoning or Memorization? Unreliable Results of Reinforcement Learning Due to Data Contamination
The reasoning capabilities of large language models (LLMs) have been a longstanding focus of research. Recent works have further enhanced these capabilities using reinforcement learning (RL), with many new methods claiming significant improvements with minimal or no external supervision. Surprisingly, some studies even suggest that random or incorrect reward signals can enhance reasoning performance. However, these breakthroughs are mostly reported on the Qwen2.5 model family and evaluated on well-known benchmarks such as MATH-500, AMC, and AIME, while failing to achieve similar gains on other models like Llama, which warrants further investigation. Our analysis shows that although Qwen2.5 achieves strong mathematical reasoning performance, its pretraining on large-scale web corpora makes it vulnerable to data contamination in popular benchmarks. As a result, results derived from these benchmarks may be unreliable. To address this, we introduce a generator that produces fully synthetic arithmetic problems of arbitrary length and difficulty, yielding a clean dataset we call RandomCalculation. Using these leakage-free datasets, we show that only accurate reward signals consistently improve performance, while noisy or incorrect signals do not. We advocate for evaluating RL methods on uncontaminated benchmarks and across diverse model families to ensure trustworthy conclusions.
comment: 26 pages
☆ Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation
Scaling language models unlocks impressive capabilities, but the accompanying computational and memory demands make both training and deployment expensive. Existing efficiency efforts typically target either parameter sharing or adaptive computation, leaving open the question of how to attain both simultaneously. We introduce Mixture-of-Recursions (MoR), a unified framework that combines the two axes of efficiency inside a single Recursive Transformer. MoR reuses a shared stack of layers across recursion steps to achieve parameter efficiency, while lightweight routers enable adaptive token-level thinking by dynamically assigning different recursion depths to individual tokens. This allows MoR to focus quadratic attention computation only among tokens still active at a given recursion depth, further improving memory access efficiency by selectively caching only their key-value pairs. Beyond these core mechanisms, we also propose a KV sharing variant that reuses KV pairs from the first recursion, specifically designed to decrease prefill latency and memory footprint. Across model scales ranging from 135M to 1.7B parameters, MoR forms a new Pareto frontier: at equal training FLOPs and smaller model sizes, it significantly lowers validation perplexity and improves few-shot accuracy, while delivering higher throughput compared with vanilla and existing recursive baselines. These gains demonstrate that MoR is an effective path towards large-model quality without incurring large-model cost.
comment: 36 pages, 9 figures, 14 tables, codes at https://github.com/raymin0223/mixture_of_recursions
☆ Benchmarking and Evaluation of AI Models in Biology: Outcomes and Recommendations from the CZI Virtual Cells Workshop
Artificial intelligence holds immense promise for transforming biology, yet a lack of standardized, cross domain, benchmarks undermines our ability to build robust, trustworthy models. Here, we present insights from a recent workshop that convened machine learning and computational biology experts across imaging, transcriptomics, proteomics, and genomics to tackle this gap. We identify major technical and systemic bottlenecks such as data heterogeneity and noise, reproducibility challenges, biases, and the fragmented ecosystem of publicly available resources and propose a set of recommendations for building benchmarking frameworks that can efficiently compare ML models of biological systems across tasks and data modalities. By promoting high quality data curation, standardized tooling, comprehensive evaluation metrics, and open, collaborative platforms, we aim to accelerate the development of robust benchmarks for AI driven Virtual Cells. These benchmarks are crucial for ensuring rigor, reproducibility, and biological relevance, and will ultimately advance the field toward integrated models that drive new discoveries, therapeutic insights, and a deeper understanding of cellular systems.
☆ National level satellite-based crop field inventories in smallholder landscapes
The design of science-based policies to improve the sustainability of smallholder agriculture is challenged by a limited understanding of fundamental system properties, such as the spatial distribution of active cropland and field size. We integrate very high spatial resolution (1.5 m) Earth observation data and deep transfer learning to derive crop field delineations in complex agricultural systems at the national scale, while maintaining minimum reference data requirements and enhancing transferability. We provide the first national-level dataset of 21 million individual fields for Mozambique (covering ~800,000 km2) for 2023. Our maps separate active cropland from non-agricultural land use with an overall accuracy of 93% and balanced omission and commission errors. Field-level spatial agreement reached median intersection over union (IoU) scores of 0.81, advancing the state-of-the-art in large-area field delineation in complex smallholder systems. The active cropland maps capture fragmented rural regions with low cropland shares not yet identified in global land cover or cropland maps. These regions are mostly located in agricultural frontier regions which host 7-9% of the Mozambican population. Field size in Mozambique is very low overall, with half of the fields being smaller than 0.16 ha, and 83% smaller than 0.5 ha. Mean field size at aggregate spatial resolution (0.05{\deg}) is 0.32 ha, but it varies strongly across gradients of accessibility, population density, and net forest cover change. This variation reflects a diverse set of actors, ranging from semi-subsistence smallholder farms to medium-scale commercial farming, and large-scale farming operations. Our results highlight that field size is a key indicator relating to socio-economic and environmental outcomes of agriculture (e.g., food production, livelihoods, deforestation, biodiversity), as well as their trade-offs.
☆ Split Happens: Combating Advanced Threats with Split Learning and Function Secret Sharing
Split Learning (SL) -- splits a model into two distinct parts to help protect client data while enhancing Machine Learning (ML) processes. Though promising, SL has proven vulnerable to different attacks, thus raising concerns about how effective it may be in terms of data privacy. Recent works have shown promising results for securing SL through the use of a novel paradigm, named Function Secret Sharing (FSS), in which servers obtain shares of a function they compute and operate on a public input hidden with a random mask. However, these works fall short in addressing the rising number of attacks which exist on SL. In SplitHappens, we expand the combination of FSS and SL to U-shaped SL. Similarly to other works, we are able to make use of the benefits of SL by reducing the communication and computational costs of FSS. However, a U-shaped SL provides a higher security guarantee than previous works, allowing a client to keep the labels of the training data secret, without having to share them with the server. Through this, we are able to generalize the security analysis of previous works and expand it to different attack vectors, such as modern model inversion attacks as well as label inference attacks. We tested our approach for two different convolutional neural networks on different datasets. These experiments show the effectiveness of our approach in reducing the training time as well as the communication costs when compared to simply using FSS while matching prior accuracy.
☆ BenchReAD: A systematic benchmark for retinal anomaly detection MICCAI 2025
Retinal anomaly detection plays a pivotal role in screening ocular and systemic diseases. Despite its significance, progress in the field has been hindered by the absence of a comprehensive and publicly available benchmark, which is essential for the fair evaluation and advancement of methodologies. Due to this limitation, previous anomaly detection work related to retinal images has been constrained by (1) a limited and overly simplistic set of anomaly types, (2) test sets that are nearly saturated, and (3) a lack of generalization evaluation, resulting in less convincing experimental setups. Furthermore, existing benchmarks in medical anomaly detection predominantly focus on one-class supervised approaches (training only with negative samples), overlooking the vast amounts of labeled abnormal data and unlabeled data that are commonly available in clinical practice. To bridge these gaps, we introduce a benchmark for retinal anomaly detection, which is comprehensive and systematic in terms of data and algorithm. Through categorizing and benchmarking previous methods, we find that a fully supervised approach leveraging disentangled representations of abnormalities (DRA) achieves the best performance but suffers from significant drops in performance when encountering certain unseen anomalies. Inspired by the memory bank mechanisms in one-class supervised learning, we propose NFM-DRA, which integrates DRA with a Normal Feature Memory to mitigate the performance degradation, establishing a new SOTA. The benchmark is publicly available at https://github.com/DopamineLcy/BenchReAD.
comment: MICCAI 2025
☆ Overcoming catastrophic forgetting in neural networks
Catastrophic forgetting is the primary challenge that hinders continual learning, which refers to a neural network ability to sequentially learn multiple tasks while retaining previously acquired knowledge. Elastic Weight Consolidation, a regularization-based approach inspired by synaptic consolidation in biological neural systems, has been used to overcome this problem. In this study prior research is replicated and extended by evaluating EWC in supervised learning settings using the PermutedMNIST and RotatedMNIST benchmarks. Through systematic comparisons with L2 regularization and stochastic gradient descent (SGD) without regularization, we analyze how different approaches balance knowledge retention and adaptability. Our results confirm what was shown in previous research, showing that EWC significantly reduces forgetting compared to naive training while slightly compromising learning efficiency on new tasks. Moreover, we investigate the impact of dropout regularization and varying hyperparameters, offering insights into the generalization of EWC across diverse learning scenarios. These results underscore EWC's potential as a viable solution for lifelong learning in neural networks.
comment: 7 pages, 5 figures, EE-411 Fundamentals of inference and learning course project
☆ The Target Polish: A New Approach to Outlier-Resistant Non-Negative Matrix and Tensor Factorization
This paper introduces the "Target Polish," a robust and computationally efficient framework for nonnegative matrix and tensor factorization. Although conventional weighted NMF approaches are resistant to outliers, they converge slowly due to the use of multiplicative updates to minimize the objective criterion. In contrast, the Target Polish approach remains compatible with the Fast-HALS algorithm, which is renowned for its speed, by adaptively smoothing the data with a weighted median-based transformation. This innovation provides outlier resistance while maintaining the highly efficient additive update structure of Fast-HALS. Empirical evaluations using image datasets corrupted with structured (block) and unstructured (salt) noise demonstrate that the Target Polish approach matches or exceeds the accuracy of state-of-the-art robust NMF methods and reduces computational time by an order of magnitude in the studied scenarios.
comment: 6 pages, 4 figures, International Conference on Robust Statistics 2025, Stresa, Italy
☆ From BERT to Qwen: Hate Detection across architectures
Online platforms struggle to curb hate speech without over-censoring legitimate discourse. Early bidirectional transformer encoders made big strides, but the arrival of ultra-large autoregressive LLMs promises deeper context-awareness. Whether this extra scale actually improves practical hate-speech detection on real-world text remains unverified. Our study puts this question to the test by benchmarking both model families, classic encoders and next-generation LLMs, on curated corpora of online interactions for hate-speech detection (Hate or No Hate).
comment: 4 pages, 5 figures. EE-559 Deep Learning course project (Group 11)
☆ RAPNet: A Receptive-Field Adaptive Convolutional Neural Network for Pansharpening
Pansharpening refers to the process of integrating a high resolution panchromatic (PAN) image with a lower resolution multispectral (MS) image to generate a fused product, which is pivotal in remote sensing. Despite the effectiveness of CNNs in addressing this challenge, they are inherently constrained by the uniform application of convolutional kernels across all spatial positions, overlooking local content variations. To overcome this issue, we introduce RAPNet, a new architecture that leverages content-adaptive convolution. At its core, RAPNet employs the Receptive-field Adaptive Pansharpening Convolution (RAPConv), designed to produce spatially adaptive kernels responsive to local feature context, thereby enhancing the precision of spatial detail extraction. Additionally, the network integrates the Pansharpening Dynamic Feature Fusion (PAN-DFF) module, which incorporates an attention mechanism to achieve an optimal balance between spatial detail enhancement and spectral fidelity. Comprehensive evaluations on publicly available datasets confirm that RAPNet delivers superior performance compared to existing approaches, as demonstrated by both quantitative metrics and qualitative assessments. Ablation analyses further substantiate the effectiveness of the proposed adaptive components.
comment: To appear in the proceedings of the 6th International Conference on Artificial Intelligence and Electromechanical Automation (AIEA 2025). 5 pages, 6 figures
☆ Logic layer Prompt Control Injection (LPCI): A Novel Security Vulnerability Class in Agentic Systems
The integration of large language models (LLMs) into enterprise systems has created a new class of covert security vulnerabilities, particularly within logic-execution layers and persistent-memory contexts. In this paper, we introduce Logic-Layer Prompt Control Injection (LPCI), a novel attack category in which encoded, delayed, and conditionally triggered payloads are embedded in memory, vector stores, or tool outputs. These payloads can bypass conventional input filters and trigger unauthorised behaviour across sessions.
☆ Some remarks on gradient dominance and LQR policy optimization
Solutions of optimization problems, including policy optimization in reinforcement learning, typically rely upon some variant of gradient descent. There has been much recent work in the machine learning, control, and optimization communities applying the Polyak-{\L}ojasiewicz Inequality (PLI) to such problems in order to establish an exponential rate of convergence (a.k.a. ``linear convergence'' in the local-iteration language of numerical analysis) of loss functions to their minima under the gradient flow. Often, as is the case of policy iteration for the continuous-time LQR problem, this rate vanishes for large initial conditions, resulting in a mixed globally linear / locally exponential behavior. This is in sharp contrast with the discrete-time LQR problem, where there is global exponential convergence. That gap between CT and DT behaviors motivates the search for various generalized PLI-like conditions, and this talk will address that topic. Moreover, these generalizations are key to understanding the transient and asymptotic effects of errors in the estimation of the gradient, errors which might arise from adversarial attacks, wrong evaluation by an oracle, early stopping of a simulation, inaccurate and very approximate digital twins, stochastic computations (algorithm ``reproducibility''), or learning by sampling from limited data. We describe an ``input to state stability'' (ISS) analysis of this issue. The lecture also discussed convergence and PLI-like properties of ``linear feedforward neural networks'' in feedback control, but this arXiv skips that part (to be updated). Much of the work described here was done in collaboration with Arthur Castello B. de Oliveira, Leilei Cui, Zhong-Ping Jiang, and Milad Siami.
comment: This is a short paper summarizing the first part of the slides presented at my keynote at the 2025 L4DC (Learning for Dynamics & Control Conference) in Ann Arbor, Michigan, 05 June 2025. A partial bibliography has been added. A second part on neural net feedback controllers is to be added
☆ CLA: Latent Alignment for Online Continual Self-Supervised Learning
Self-supervised learning (SSL) is able to build latent representations that generalize well to unseen data. However, only a few SSL techniques exist for the online CL setting, where data arrives in small minibatches, the model must comply with a fixed computational budget, and task boundaries are absent. We introduce Continual Latent Alignment (CLA), a novel SSL strategy for Online CL that aligns the representations learned by the current model with past representations to mitigate forgetting. We found that our CLA is able to speed up the convergence of the training process in the online scenario, outperforming state-of-the-art approaches under the same computational budget. Surprisingly, we also discovered that using CLA as a pretraining protocol in the early stages of pretraining leads to a better final performance when compared to a full i.i.d. pretraining.
comment: Accepted at CoLLAs 2025 conference
☆ Efficient Federated Learning with Heterogeneous Data and Adaptive Dropout KDD
Federated Learning (FL) is a promising distributed machine learning approach that enables collaborative training of a global model using multiple edge devices. The data distributed among the edge devices is highly heterogeneous. Thus, FL faces the challenge of data distribution and heterogeneity, where non-Independent and Identically Distributed (non-IID) data across edge devices may yield in significant accuracy drop. Furthermore, the limited computation and communication capabilities of edge devices increase the likelihood of stragglers, thus leading to slow model convergence. In this paper, we propose the FedDHAD FL framework, which comes with two novel methods: Dynamic Heterogeneous model aggregation (FedDH) and Adaptive Dropout (FedAD). FedDH dynamically adjusts the weights of each local model within the model aggregation process based on the non-IID degree of heterogeneous data to deal with the statistical data heterogeneity. FedAD performs neuron-adaptive operations in response to heterogeneous devices to improve accuracy while achieving superb efficiency. The combination of these two methods makes FedDHAD significantly outperform state-of-the-art solutions in terms of accuracy (up to 6.7% higher), efficiency (up to 2.02 times faster), and computation cost (up to 15.0% smaller).
comment: 29 pages, to appear in ACM Transactions on Knowledge Discovery from Data (TKDD)
☆ Non-exchangeable Conformal Prediction with Optimal Transport: Tackling Distribution Shifts with Unlabeled Data
Conformal prediction is a distribution-free uncertainty quantification method that has gained popularity in the machine learning community due to its finite-sample guarantees and ease of use. Its most common variant, dubbed split conformal prediction, is also computationally efficient as it boils down to collecting statistics of the model predictions on some calibration data not yet seen by the model. Nonetheless, these guarantees only hold if the calibration and test data are exchangeable, a condition that is difficult to verify and often violated in practice due to so-called distribution shifts. The literature is rife with methods to mitigate the loss in coverage in this non-exchangeable setting, but these methods require some prior information on the type of distribution shift to be expected at test time. In this work, we study this problem via a new perspective, through the lens of optimal transport, and show that it is possible to estimate the loss in coverage and mitigate it in case of distribution shift.
☆ SentiDrop: A Multi Modal Machine Learning model for Predicting Dropout in Distance Learning
School dropout is a serious problem in distance learning, where early detection is crucial for effective intervention and student perseverance. Predicting student dropout using available educational data is a widely researched topic in learning analytics. Our partner's distance learning platform highlights the importance of integrating diverse data sources, including socio-demographic data, behavioral data, and sentiment analysis, to accurately predict dropout risks. In this paper, we introduce a novel model that combines sentiment analysis of student comments using the Bidirectional Encoder Representations from Transformers (BERT) model with socio-demographic and behavioral data analyzed through Extreme Gradient Boosting (XGBoost). We fine-tuned BERT on student comments to capture nuanced sentiments, which were then merged with key features selected using feature importance techniques in XGBoost. Our model was tested on unseen data from the next academic year, achieving an accuracy of 84\%, compared to 82\% for the baseline model. Additionally, the model demonstrated superior performance in other metrics, such as precision and F1-score. The proposed method could be a vital tool in developing personalized strategies to reduce dropout rates and encourage student perseverance
comment: International Conference on Education and New Learning Technologies (2025)
☆ Multiple Choice Learning of Low Rank Adapters for Language Modeling
We propose LoRA-MCL, a training scheme that extends next-token prediction in language models with a method designed to decode diverse, plausible sentence continuations at inference time. Traditional language modeling is an intrinsically ill-posed problem: given a context, multiple futures may be equally plausible. Our approach leverages Multiple Choice Learning (MCL) and the Winner-Takes-All (WTA) loss to efficiently handle ambiguity through Low-Rank Adaptation (LoRA). We provide a theoretical interpretation of applying Multiple Choice Learning to Language Modeling, assuming the data is generated from a mixture of distributions. To illustrate the proposed approach, we use data sampled from mixtures of Markov chains. We then demonstrate with extensive experiments on real-world visual and audio captioning tasks that our method achieves high diversity and relevance in generated outputs.
☆ Energy Efficiency in AI for 5G and Beyond: A DeepRx Case Study
This study addresses the challenge of balancing energy efficiency with performance in AI/ML models, focusing on DeepRX, a deep learning receiver based on a fully convolutional ResNet architecture. We evaluate the energy consumption of DeepRX, considering factors including FLOPs/Watt and FLOPs/clock, and find consistency between estimated and actual energy usage, influenced by memory access patterns. The research extends to comparing energy dynamics during training and inference phases. A key contribution is the application of knowledge distillation (KD) to train a compact DeepRX \textit{student} model that emulates the performance of the \textit{teacher} model but with reduced energy consumption. We experiment with different student model sizes, optimal teacher sizes, and KD hyperparameters. Performance is measured by comparing the Bit Error Rate (BER) performance versus Signal-to-Interference \& Noise Ratio (SINR) values of the distilled model and a model trained from scratch. The distilled models demonstrate a lower error floor across SINR levels, highlighting the effectiveness of KD in achieving energy-efficient AI solutions.
☆ Anticipating the Selectivity of Cyclization Reaction Pathways with Neural Network Potentials
Reaction mechanism search tools have demonstrated the ability to provide insights into likely products and rate-limiting steps of reacting systems. However, reactions involving several concerted bond changes - as can be found in many key steps of natural product synthesis - can complicate the search process. To mitigate these complications, we present a mechanism search strategy particularly suited to help expedite exploration of an exemplary family of such complex reactions, cyclizations. We provide a cost-effective strategy for identifying relevant elementary reaction steps by combining graph-based enumeration schemes and machine learning techniques for intermediate filtering. Key to this approach is our use of a neural network potential (NNP), AIMNet2-rxn, for computational evaluation of each candidate reaction pathway. In this article, we evaluate the NNP's ability to estimate activation energies, demonstrate the correct anticipation of stereoselectivity, and recapitulate complex enabling steps in natural product synthesis.
comment: 32 pages, 5 figures
☆ Extracting Important Tokens in E-Commerce Queries with a Tag Interaction-Aware Transformer Model
The major task of any e-commerce search engine is to retrieve the most relevant inventory items, which best match the user intent reflected in a query. This task is non-trivial due to many reasons, including ambiguous queries, misaligned vocabulary between buyers, and sellers, over- or under-constrained queries by the presence of too many or too few tokens. To address these challenges, query reformulation is used, which modifies a user query through token dropping, replacement or expansion, with the objective to bridge semantic gap between query tokens and users' search intent. Early methods of query reformulation mostly used statistical measures derived from token co-occurrence frequencies from selective user sessions having clicks or purchases. In recent years, supervised deep learning approaches, specifically transformer-based neural language models, or sequence-to-sequence models are being used for query reformulation task. However, these models do not utilize the semantic tags of a query token, which are significant for capturing user intent of an e-commerce query. In this work, we pose query reformulation as a token classification task, and solve this task by designing a dependency-aware transformer-based language model, TagBERT, which makes use of semantic tags of a token for learning superior query phrase embedding. Experiments on large, real-life e-commerce datasets show that TagBERT exhibits superior performance than plethora of competing models, including BERT, eBERT, and Sequence-to-Sequence transformer model for important token classification task.
☆ Dynamical stability for dense patterns in discrete attractor neural networks
Neural networks storing multiple discrete attractors are canonical models of biological memory. Previously, the dynamical stability of such networks could only be guaranteed under highly restrictive conditions. Here, we derive a theory of the local stability of discrete fixed points in a broad class of networks with graded neural activities and in the presence of noise. By directly analyzing the bulk and outliers of the Jacobian spectrum, we show that all fixed points are stable below a critical load that is distinct from the classical \textit{critical capacity} and depends on the statistics of neural activities in the fixed points as well as the single-neuron activation function. Our analysis highlights the computational benefits of threshold-linear activation and sparse-like patterns.
☆ Leveraging RAG-LLMs for Urban Mobility Simulation and Analysis
With the rise of smart mobility and shared e-mobility services, numerous advanced technologies have been applied to this field. Cloud-based traffic simulation solutions have flourished, offering increasingly realistic representations of the evolving mobility landscape. LLMs have emerged as pioneering tools, providing robust support for various applications, including intelligent decision-making, user interaction, and real-time traffic analysis. As user demand for e-mobility continues to grow, delivering comprehensive end-to-end solutions has become crucial. In this paper, we present a cloud-based, LLM-powered shared e-mobility platform, integrated with a mobile application for personalized route recommendations. The optimization module is evaluated based on travel time and cost across different traffic scenarios. Additionally, the LLM-powered RAG framework is evaluated at the schema level for different users, using various evaluation methods. Schema-level RAG with XiYanSQL achieves an average execution accuracy of 0.81 on system operator queries and 0.98 on user queries.
☆ Improving Remote Sensing Classification using Topological Data Analysis and Convolutional Neural Networks
Topological data analysis (TDA) is a relatively new field that is gaining rapid adoption due to its robustness and ability to effectively describe complex datasets by quantifying geometric information. In imaging contexts, TDA typically models data as filtered cubical complexes from which we can extract discriminative features using persistence homology. Meanwhile, convolutional neural networks (CNNs) have been shown to be biased towards texture based local features. To address this limitation, we propose a TDA feature engineering pipeline and a simple method to integrate topological features with deep learning models on remote sensing classification. Our method improves the performance of a ResNet18 model on the EuroSAT dataset by 1.44% achieving 99.33% accuracy, which surpasses all previously reported single-model accuracies, including those with larger architectures, such as ResNet50 (2x larger) and XL Vision Transformers (197x larger). We additionally show that our method's accuracy is 1.82% higher than our ResNet18 baseline on the RESISC45 dataset. To our knowledge, this is the first application of TDA features in satellite scene classification with deep learning. This demonstrates that TDA features can be integrated with deep learning models, even on datasets without explicit topological structures, thereby increasing the applicability of TDA. A clean implementation of our method will be made publicly available upon publication.
comment: 9 pages, 8 figures
☆ Test-Time Canonicalization by Foundation Models for Robust Perception ICML 2025
Real-world visual perception requires invariance to diverse transformations, yet current methods rely heavily on specialized architectures or training on predefined augmentations, limiting generalization. We propose FOCAL, a test-time, data-driven framework that achieves robust perception by leveraging internet-scale visual priors from foundation models. By generating and optimizing candidate transformations toward visually typical, "canonical" views, FOCAL enhances robustness without re-training or architectural changes. Our experiments demonstrate improved robustness of CLIP and SAM across challenging transformations, including 2D/3D rotations, illumination shifts (contrast and color), and day-night variations. We also highlight potential applications in active vision. Our approach challenges the assumption that transform-specific training is necessary, instead offering a scalable path to invariance. Our code is available at: https://github.com/sutkarsh/focal.
comment: Published at ICML 2025
☆ Enhanced DeepONet for 1-D consolidation operator learning: an architectural investigation
Deep Operator Networks (DeepONets) have emerged as a powerful surrogate modeling framework for learning solution operators in PDE-governed systems. While their use is expanding across engineering disciplines, applications in geotechnical engineering remain limited. This study systematically evaluates several DeepONet architectures for the one-dimensional consolidation problem. We initially consider three architectures: a standard DeepONet with the coefficient of consolidation embedded in the branch net (Models 1 and 2), and a physics-inspired architecture with the coefficient embedded in the trunk net (Model 3). Results show that Model 3 outperforms the standard configurations (Models 1 and 2) but still has limitations when the target solution (excess pore pressures) exhibits significant variation. To overcome this limitation, we propose a Trunknet Fourier feature-enhanced DeepONet (Model 4) that addresses the identified limitations by capturing rapidly varying functions. All proposed architectures achieve speedups ranging from 1.5 to 100 times over traditional explicit and implicit solvers, with Model 4 being the most efficient. Larger computational savings are expected for more complex systems than the explored 1D case, which is promising. Overall, the study highlights the potential of DeepONets to enable efficient, generalizable surrogate modeling in geotechnical applications, advancing the integration of scientific machine learning in geotechnics, which is at an early stage.
☆ TAT: Temporal-Aligned Transformer for Multi-Horizon Peak Demand Forecasting KDD 2025
Multi-horizon time series forecasting has many practical applications such as demand forecasting. Accurate demand prediction is critical to help make buying and inventory decisions for supply chain management of e-commerce and physical retailers, and such predictions are typically required for future horizons extending tens of weeks. This is especially challenging during high-stake sales events when demand peaks are particularly difficult to predict accurately. However, these events are important not only for managing supply chain operations but also for ensuring a seamless shopping experience for customers. To address this challenge, we propose Temporal-Aligned Transformer (TAT), a multi-horizon forecaster leveraging apriori-known context variables such as holiday and promotion events information for improving predictive performance. Our model consists of an encoder and decoder, both embedded with a novel Temporal Alignment Attention (TAA), designed to learn context-dependent alignment for peak demand forecasting. We conduct extensive empirical analysis on two large-scale proprietary datasets from a large e-commerce retailer. We demonstrate that TAT brings up to 30% accuracy improvement on peak demand forecasting while maintaining competitive overall performance compared to other state-of-the-art methods.
comment: 9 pages, 4 figures, 7 tables, published at KDD 2025 workshop on AI for Supply Chain: Today and Future
☆ Feature Distillation is the Better Choice for Model-Heterogeneous Federated Learning
Model-Heterogeneous Federated Learning (Hetero-FL) has attracted growing attention for its ability to aggregate knowledge from heterogeneous models while keeping private data locally. To better aggregate knowledge from clients, ensemble distillation, as a widely used and effective technique, is often employed after global aggregation to enhance the performance of the global model. However, simply combining Hetero-FL and ensemble distillation does not always yield promising results and can make the training process unstable. The reason is that existing methods primarily focus on logit distillation, which, while being model-agnostic with softmax predictions, fails to compensate for the knowledge bias arising from heterogeneous models. To tackle this challenge, we propose a stable and efficient Feature Distillation for model-heterogeneous Federated learning, dubbed FedFD, that can incorporate aligned feature information via orthogonal projection to integrate knowledge from heterogeneous models better. Specifically, a new feature-based ensemble federated knowledge distillation paradigm is proposed. The global model on the server needs to maintain a projection layer for each client-side model architecture to align the features separately. Orthogonal techniques are employed to re-parameterize the projection layer to mitigate knowledge bias from heterogeneous models and thus maximize the distilled knowledge. Extensive experiments show that FedFD achieves superior performance compared to state-of-the-art methods.
☆ Parallel Sampling of Diffusion Models on $SO(3)$
In this paper, we design an algorithm to accelerate the diffusion process on the $SO(3)$ manifold. The inherently sequential nature of diffusion models necessitates substantial time for denoising perturbed data. To overcome this limitation, we proposed to adapt the numerical Picard iteration for the $SO(3)$ space. We demonstrate our algorithm on an existing method that employs diffusion models to address the pose ambiguity problem. Moreover, we show that this acceleration advantage occurs without any measurable degradation in task reward. The experiments reveal that our algorithm achieves a speed-up of up to 4.9$\times$, significantly reducing the latency for generating a single sample.
comment: MVA2025
☆ Some Super-approximation Rates of ReLU Neural Networks for Korobov Functions
This paper examines the $L_p$ and $W^1_p$ norm approximation errors of ReLU neural networks for Korobov functions. In terms of network width and depth, we derive nearly optimal super-approximation error bounds of order $2m$ in the $L_p$ norm and order $2m-2$ in the $W^1_p$ norm, for target functions with $L_p$ mixed derivative of order $m$ in each direction. The analysis leverages sparse grid finite elements and the bit extraction technique. Our results improve upon classical lowest order $L_\infty$ and $H^1$ norm error bounds and demonstrate that the expressivity of neural networks is largely unaffected by the curse of dimensionality.
☆ MoCap-Impute: A Comprehensive Benchmark and Comparative Analysis of Imputation Methods for IMU-based Motion Capture Data
Motion capture (MoCap) data from wearable Inertial Measurement Units (IMUs) is vital for applications in sports science, but its utility is often compromised by missing data. Despite numerous imputation techniques, a systematic performance evaluation for IMU-derived MoCap time-series data is lacking. We address this gap by conducting a comprehensive comparative analysis of statistical, machine learning, and deep learning imputation methods. Our evaluation considers three distinct contexts: univariate time-series, multivariate across subjects, and multivariate across kinematic angles. To facilitate this benchmark, we introduce the first publicly available MoCap dataset designed specifically for imputation, featuring data from 53 karate practitioners. We simulate three controlled missingness mechanisms: missing completely at random (MCAR), block missingness, and a novel value-dependent pattern at signal transition points. Our experiments, conducted on 39 kinematic variables across all subjects, reveal that multivariate imputation frameworks consistently outperform univariate approaches, particularly for complex missingness. For instance, multivariate methods achieve up to a 50% mean absolute error reduction (MAE from 10.8 to 5.8) compared to univariate techniques for transition point missingness. Advanced models like Generative Adversarial Imputation Networks (GAIN) and Iterative Imputers demonstrate the highest accuracy in these challenging scenarios. This work provides a critical baseline for future research and offers practical recommendations for improving the integrity and robustness of Mo-Cap data analysis.
comment: 22 pages, 7 figures, 3 algorithms, 2 tables
☆ Bridging Robustness and Generalization Against Word Substitution Attacks in NLP via the Growth Bound Matrix Approach ACL
Despite advancements in Natural Language Processing (NLP), models remain vulnerable to adversarial attacks, such as synonym substitutions. While prior work has focused on improving robustness for feed-forward and convolutional architectures, the robustness of recurrent networks and modern state space models (SSMs), such as S4, remains understudied. These architectures pose unique challenges due to their sequential processing and complex parameter dynamics. In this paper, we introduce a novel regularization technique based on Growth Bound Matrices (GBM) to improve NLP model robustness by reducing the impact of input perturbations on model outputs. We focus on computing the GBM for three architectures: Long Short-Term Memory (LSTM), State Space models (S4), and Convolutional Neural Networks (CNN). Our method aims to (1) enhance resilience against word substitution attacks, (2) improve generalization on clean text, and (3) providing the first systematic analysis of SSM (S4) robustness. Extensive experiments across multiple architectures and benchmark datasets demonstrate that our method improves adversarial robustness by up to 8.8% over existing baselines. These results highlight the effectiveness of our approach, outperforming several state-of-the-art methods in adversarial defense. Codes are available at https://github.com/BouriMohammed/GBM
comment: Accepted to ACL Findings 2025
☆ Convergence of Agnostic Federated Averaging
Federated learning (FL) enables decentralized model training without centralizing raw data. However, practical FL deployments often face a key realistic challenge: Clients participate intermittently in server aggregation and with unknown, possibly biased participation probabilities. Most existing convergence results either assume full-device participation, or rely on knowledge of (in fact uniform) client availability distributions -- assumptions that rarely hold in practice. In this work, we characterize the optimization problem that consistently adheres to the stochastic dynamics of the well-known \emph{agnostic Federated Averaging (FedAvg)} algorithm under random (and variably-sized) client availability, and rigorously establish its convergence for convex, possibly nonsmooth losses, achieving a standard rate of order $\mathcal{O}(1/\sqrt{T})$, where $T$ denotes the aggregation horizon. Our analysis provides the first convergence guarantees for agnostic FedAvg under general, non-uniform, stochastic client participation, without knowledge of the participation distribution. We also empirically demonstrate that agnostic FedAvg in fact outperforms common (and suboptimal) weighted aggregation FedAvg variants, even with server-side knowledge of participation weights.
comment: 5 pages, 2 figurres, CAMSAP conference
☆ Recognizing Dementia from Neuropsychological Tests with State Space Models
Early detection of dementia is critical for timely medical intervention and improved patient outcomes. Neuropsychological tests are widely used for cognitive assessment but have traditionally relied on manual scoring. Automatic dementia classification (ADC) systems aim to infer cognitive decline directly from speech recordings of such tests. We propose Demenba, a novel ADC framework based on state space models, which scale linearly in memory and computation with sequence length. Trained on over 1,000 hours of cognitive assessments administered to Framingham Heart Study participants, some of whom were diagnosed with dementia through adjudicated review, our method outperforms prior approaches in fine-grained dementia classification by 21\%, while using fewer parameters. We further analyze its scaling behavior and demonstrate that our model gains additional improvement when fused with large language models, paving the way for more transparent and scalable dementia assessment tools. Code: https://anonymous.4open.science/r/Demenba-0861
☆ MF-GLaM: A multifidelity stochastic emulator using generalized lambda models
Stochastic simulators exhibit intrinsic stochasticity due to unobservable, uncontrollable, or unmodeled input variables, resulting in random outputs even at fixed input conditions. Such simulators are common across various scientific disciplines; however, emulating their entire conditional probability distribution is challenging, as it is a task traditional deterministic surrogate modeling techniques are not designed for. Additionally, accurately characterizing the response distribution can require prohibitively large datasets, especially for computationally expensive high-fidelity (HF) simulators. When lower-fidelity (LF) stochastic simulators are available, they can enhance limited HF information within a multifidelity surrogate modeling (MFSM) framework. While MFSM techniques are well-established for deterministic settings, constructing multifidelity emulators to predict the full conditional response distribution of stochastic simulators remains a challenge. In this paper, we propose multifidelity generalized lambda models (MF-GLaMs) to efficiently emulate the conditional response distribution of HF stochastic simulators by exploiting data from LF stochastic simulators. Our approach builds upon the generalized lambda model (GLaM), which represents the conditional distribution at each input by a flexible, four-parameter generalized lambda distribution. MF-GLaMs are non-intrusive, requiring no access to the internal stochasticity of the simulators nor multiple replications of the same input values. We demonstrate the efficacy of MF-GLaM through synthetic examples of increasing complexity and a realistic earthquake application. Results show that MF-GLaMs can achieve improved accuracy at the same cost as single-fidelity GLaMs, or comparable performance at significantly reduced cost.
☆ Average Sensitivity of Hierarchical $k$-Median Clustering
Hierarchical clustering is a widely used method for unsupervised learning with numerous applications. However, in the application of modern algorithms, the datasets studied are usually large and dynamic. If the hierarchical clustering is sensitive to small perturbations of the dataset, the usability of the algorithm will be greatly reduced. In this paper, we focus on the hierarchical $k$ -median clustering problem, which bridges hierarchical and centroid-based clustering while offering theoretical appeal, practical utility, and improved interpretability. We analyze the average sensitivity of algorithms for this problem by measuring the expected change in the output when a random data point is deleted. We propose an efficient algorithm for hierarchical $k$-median clustering and theoretically prove its low average sensitivity and high clustering quality. Additionally, we show that single linkage clustering and a deterministic variant of the CLNSS algorithm exhibit high average sensitivity, making them less stable. Finally, we validate the robustness and effectiveness of our algorithm through experiments.
☆ Conditional Chemical Language Models are Versatile Tools in Drug Discovery
Generative chemical language models (CLMs) have demonstrated strong capabilities in molecular design, yet their impact in drug discovery remains limited by the absence of reliable reward signals and the lack of interpretability in their outputs. We present SAFE-T, a generalist chemical modeling framework that conditions on biological context -- such as protein targets or mechanisms of action -- to prioritize and design molecules without relying on structural information or engineered scoring functions. SAFE-T models the conditional likelihood of fragment-based molecular sequences given a biological prompt, enabling principled scoring of molecules across tasks such as virtual screening, drug-target interaction prediction, and activity cliff detection. Moreover, it supports goal-directed generation by sampling from this learned distribution, aligning molecular design with biological objectives. In comprehensive zero-shot evaluations across predictive (LIT-PCBA, DAVIS, KIBA, ACNet) and generative (DRUG, PMO) benchmarks, SAFE-T consistently achieves performance comparable to or better than existing approaches while being significantly faster. Fragment-level attribution further reveals that SAFE-T captures known structure-activity relationships, supporting interpretable and biologically grounded design. Together with its computational efficiency, these results demonstrate that conditional generative CLMs can unify scoring and generation to accelerate early-stage drug discovery.
comment: 12 pages, extra 13 pages of appendix
☆ DNS Tunneling: Threat Landscape and Improved Detection Solutions
Detecting Domain Name System (DNS) tunneling is a significant challenge in security due to its capacity to hide harmful actions within DNS traffic that appears to be normal and legitimate. Traditional detection methods are based on rule-based approaches or signature matching methods that are often insufficient to accurately identify such covert communication channels. This research is about effectively detecting DNS tunneling. We propose a novel approach to detect DNS tunneling with machine learning algorithms. We combine machine learning algorithms to analyze the traffic by using features extracted from DNS traffic. Analyses results show that the proposed approach is a good candidate to detect DNS tunneling accurately.
☆ DepViT-CAD: Deployable Vision Transformer-Based Cancer Diagnosis in Histopathology
Accurate and timely cancer diagnosis from histopathological slides is vital for effective clinical decision-making. This paper introduces DepViT-CAD, a deployable AI system for multi-class cancer diagnosis in histopathology. At its core is MAViT, a novel Multi-Attention Vision Transformer designed to capture fine-grained morphological patterns across diverse tumor types. MAViT was trained on expert-annotated patches from 1008 whole-slide images, covering 11 diagnostic categories, including 10 major cancers and non-tumor tissue. DepViT-CAD was validated on two independent cohorts: 275 WSIs from The Cancer Genome Atlas and 50 routine clinical cases from pathology labs, achieving diagnostic sensitivities of 94.11% and 92%, respectively. By combining state-of-the-art transformer architecture with large-scale real-world validation, DepViT-CAD offers a robust and scalable approach for AI-assisted cancer diagnostics. To support transparency and reproducibility, software and code will be made publicly available at GitHub.
comment: 25 pages, 15 figures
☆ Kernel-Adaptive PI-ELMs for Forward and Inverse Problems in PDEs with Sharp Gradients
This paper introduces the Kernel Adaptive Physics-Informed Extreme Learning Machine (KAPI-ELM), an adaptive Radial Basis Function (RBF)-based extension of PI-ELM designed to solve both forward and inverse Partial Differential Equation (PDE) problems involving localized sharp gradients. While PI-ELMs outperform the traditional Physics-Informed Neural Networks (PINNs) in speed due to their single-shot, least square optimization, this advantage comes at a cost: their fixed, randomly initialized input layer limits their ability to capture sharp gradients. To overcome this limitation, we introduce a lightweight Bayesian Optimization (BO) framework that, instead of adjusting each input layer parameter individually as in traditional backpropagation, learns a small set of hyperparameters defining the statistical distribution from which the input weights are drawn. This novel distributional optimization strategy -- combining BO for input layer distributional parameters with least-squares optimization for output layer network parameters -- enables KAPI-ELM to preserve PI-ELM's speed while matching or exceeding the expressiveness of PINNs. We validate the proposed methodology on several challenging forward and inverse PDE benchmarks, including a 1D singularly perturbed convection-diffusion equation, a 2D Poisson equation with sharp localized sources, and a time-dependent advection equation. Notably, KAPI-ELM achieves state-of-the-art accuracy in both forward and inverse settings. In stiff PDE regimes, it matches or even outperforms advanced methods such as the Extended Theory of Functional Connections (XTFC), while requiring nearly an order of magnitude fewer tunable parameters. These results establish the potential of KAPI-ELM as a scalable, interpretable, and generalizable physics-informed learning framework, especially in stiff PDE regimes.
☆ Visual Analytics for Explainable and Trustworthy Artificial Intelligence
Our society increasingly depends on intelligent systems to solve complex problems, ranging from recommender systems suggesting the next movie to watch to AI models assisting in medical diagnoses for hospitalized patients. With the iterative improvement of diagnostic accuracy and efficiency, AI holds significant potential to mitigate medical misdiagnoses by preventing numerous deaths and reducing an economic burden of approximately 450 EUR billion annually. However, a key obstacle to AI adoption lies in the lack of transparency: many automated systems function as "black boxes," providing predictions without revealing the underlying processes. This opacity can hinder experts' ability to trust and rely on AI systems. Visual analytics (VA) provides a compelling solution by combining AI models with interactive visualizations. These specialized charts and graphs empower users to incorporate their domain expertise to refine and improve the models, bridging the gap between AI and human understanding. In this work, we define, categorize, and explore how VA solutions can foster trust across the stages of a typical AI pipeline. We propose a design space for innovative visualizations and present an overview of our previously developed VA dashboards, which support critical tasks within the various pipeline stages, including data processing, feature engineering, hyperparameter tuning, understanding, debugging, refining, and comparing models.
☆ Spatial Lifting for Dense Prediction
We present Spatial Lifting (SL), a novel methodology for dense prediction tasks. SL operates by lifting standard inputs, such as 2D images, into a higher-dimensional space and subsequently processing them using networks designed for that higher dimension, such as a 3D U-Net. Counterintuitively, this dimensionality lifting allows us to achieve good performance on benchmark tasks compared to conventional approaches, while reducing inference costs and significantly lowering the number of model parameters. The SL framework produces intrinsically structured outputs along the lifted dimension. This emergent structure facilitates dense supervision during training and enables robust, near-zero-additional-cost prediction quality assessment at test time. We validate our approach across 19 benchmark datasets (13 for semantic segmentation and 6 for depth estimation), demonstrating competitive dense prediction performance while reducing the model parameter count by over 98% (in the U-Net case) and lowering inference costs. Spatial Lifting introduces a new vision modeling paradigm that offers a promising path toward more efficient, accurate, and reliable deep networks for dense prediction tasks in vision.
comment: Preprint. Under review
☆ A Graph Sufficiency Perspective for Neural Networks
This paper analyzes neural networks through graph variables and statistical sufficiency. We interpret neural network layers as graph-based transformations, where neurons act as pairwise functions between inputs and learned anchor points. Within this formulation, we establish conditions under which layer outputs are sufficient for the layer inputs, that is, each layer preserves the conditional distribution of the target variable given the input variable. Under dense anchor point assumptions, we prove that asymptotic sufficiency holds in the infinite-width limit and is preserved throughout training. To align more closely with practical architectures, we further show that sufficiency can be achieved with finite-width networks by assuming region-separated input distributions and constructing appropriate anchor points. Our framework covers fully connected layers, general pairwise functions, ReLU and sigmoid activations, and convolutional neural networks. This work bridges statistical sufficiency, graph-theoretic representations, and deep learning, providing a new statistical understanding of neural networks.
comment: 23 pages
☆ History Matching under Uncertainty of Geological Scenarios with Implicit Geological Realism Control with Generative Deep Learning and Graph Convolutions
The graph-based variational autoencoder represents an architecture that can handle the uncertainty of different geological scenarios, such as depositional or structural, through the concept of a lowerdimensional latent space. The main difference from recent studies is utilisation of a graph-based approach in reservoir modelling instead of the more traditional lattice-based deep learning methods. We provide a solution to implicitly control the geological realism through the latent variables of a generative model and Geodesic metrics. Our experiments of AHM with synthetic dataset that consists of 3D realisations of channelised geological representations with two distinct scenarios with one and two channels shows the viability of the approach. We offer in-depth analysis of the latent space using tools such as PCA, t-SNE, and TDA to illustrate its structure.
comment: Part of the completed PhD thesis https://geodatascience.hw.ac.uk/theses/
☆ Learning Private Representations through Entropy-based Adversarial Training
How can we learn a representation with high predictive power while preserving user privacy? We present an adversarial representation learning method for sanitizing sensitive content from the learned representation. Specifically, we introduce a variant of entropy - focal entropy, which mitigates the potential information leakage of the existing entropy-based approaches. We showcase feasibility on multiple benchmarks. The results suggest high target utility at moderate privacy leakage.
☆ T-GRAB: A Synthetic Diagnostic Benchmark for Learning on Temporal Graphs KDD 2025
Dynamic graph learning methods have recently emerged as powerful tools for modelling relational data evolving through time. However, despite extensive benchmarking efforts, it remains unclear whether current Temporal Graph Neural Networks (TGNNs) effectively capture core temporal patterns such as periodicity, cause-and-effect, and long-range dependencies. In this work, we introduce the Temporal Graph Reasoning Benchmark (T-GRAB), a comprehensive set of synthetic tasks designed to systematically probe the capabilities of TGNNs to reason across time. T-GRAB provides controlled, interpretable tasks that isolate key temporal skills: counting/memorizing periodic repetitions, inferring delayed causal effects, and capturing long-range dependencies over both spatial and temporal dimensions. We evaluate 11 temporal graph learning methods on these tasks, revealing fundamental shortcomings in their ability to generalize temporal patterns. Our findings offer actionable insights into the limitations of current models, highlight challenges hidden by traditional real-world benchmarks, and motivate the development of architectures with stronger temporal reasoning abilities. The code for T-GRAB can be found at: https://github.com/alirezadizaji/T-GRAB.
comment: Accepted to MLoG-GenAI Workshop @ KDD 2025 (Oral)
☆ Pimba: A Processing-in-Memory Acceleration for Post-Transformer Large Language Model Serving
Transformers are the driving force behind today's Large Language Models (LLMs), serving as the foundation for their performance and versatility. Yet, their compute and memory costs grow with sequence length, posing scalability challenges for long-context inferencing. In response, the algorithm community is exploring alternative architectures, such as state space models (SSMs), linear attention, and recurrent neural networks (RNNs), which we refer to as post-transformers. This shift presents a key challenge: building a serving system that efficiently supports both transformer and post-transformer LLMs within a unified framework. To address this challenge, we analyze the performance characteristics of transformer and post-transformer LLMs. Despite their algorithmic differences, both are fundamentally limited by memory bandwidth under batched inference due to attention in transformers and state updates in post-transformers. Further analyses suggest two additional insights: (1) state update operations, unlike attention, incur high hardware cost, making per-bank PIM acceleration inefficient, and (2) different low-precision arithmetic methods offer varying accuracy-area tradeoffs, while we identify Microsoft's MX as the Pareto-optimal choice. Building on these insights, we design Pimba as an array of State-update Processing Units (SPUs), each shared between two banks to enable interleaved access to PIM. Each SPU includes a State-update Processing Engine (SPE) that comprises element-wise multipliers and adders using MX-based quantized arithmetic, enabling efficient execution of state update and attention operations. Our evaluation shows that, compared to LLM-optimized GPU and GPU+PIM systems, Pimba achieves up to 3.2x and 2.1x higher token generation throughput, respectively.
☆ Should We Ever Prefer Decision Transformer for Offline Reinforcement Learning?
In recent years, extensive work has explored the application of the Transformer architecture to reinforcement learning problems. Among these, Decision Transformer (DT) has gained particular attention in the context of offline reinforcement learning due to its ability to frame return-conditioned policy learning as a sequence modeling task. Most recently, Bhargava et al. (2024) provided a systematic comparison of DT with more conventional MLP-based offline RL algorithms, including Behavior Cloning (BC) and Conservative Q-Learning (CQL), and claimed that DT exhibits superior performance in sparse-reward and low-quality data settings. In this paper, through experimentation on robotic manipulation tasks (Robomimic) and locomotion benchmarks (D4RL), we show that MLP-based Filtered Behavior Cloning (FBC) achieves competitive or superior performance compared to DT in sparse-reward environments. FBC simply filters out low-performing trajectories from the dataset and then performs ordinary behavior cloning on the filtered dataset. FBC is not only very straightforward, but it also requires less training data and is computationally more efficient. The results therefore suggest that DT is not preferable for sparse-reward environments. From prior work, arguably, DT is also not preferable for dense-reward environments. Thus, we pose the question: Is DT ever preferable?
comment: Accepted by RLBrew: Ingredients for Developing Generalist Agents workshop (RLC 2025)
☆ Play Style Identification Using Low-Level Representations of Play Traces in MicroRTS
Play style identification can provide valuable game design insights and enable adaptive experiences, with the potential to improve game playing agents. Previous work relies on domain knowledge to construct play trace representations using handcrafted features. More recent approaches incorporate the sequential structure of play traces but still require some level of domain abstraction. In this study, we explore the use of unsupervised CNN-LSTM autoencoder models to obtain latent representations directly from low-level play trace data in MicroRTS. We demonstrate that this approach yields a meaningful separation of different game playing agents in the latent space, reducing reliance on domain expertise and its associated biases. This latent space is then used to guide the exploration of diverse play styles within studied AI players.
comment: Accepted as Short Paper for IEEE CoG
☆ Understanding the Rank of Tensor Networks via an Intuitive Example-Driven Approach
Tensor Network (TN) decompositions have emerged as an indispensable tool in Big Data analytics owing to their ability to provide compact low-rank representations, thus alleviating the ``Curse of Dimensionality'' inherent in handling higher-order data. At the heart of their success lies the concept of TN ranks, which governs the efficiency and expressivity of TN decompositions. However, unlike matrix ranks, TN ranks often lack a universal meaning and an intuitive interpretation, with their properties varying significantly across different TN structures. Consequently, TN ranks are frequently treated as empirically tuned hyperparameters, rather than as key design parameters inferred from domain knowledge. The aim of this Lecture Note is therefore to demystify the foundational yet frequently misunderstood concept of TN ranks through real-life examples and intuitive visualizations. We begin by illustrating how domain knowledge can guide the selection of TN ranks in widely-used models such as the Canonical Polyadic (CP) and Tucker decompositions. For more complex TN structures, we employ a self-explanatory graphical approach that generalizes to tensors of arbitrary order. Such a perspective naturally reveals the relationship between TN ranks and the corresponding ranks of tensor unfoldings (matrices), thereby circumventing cumbersome multi-index tensor algebra while facilitating domain-informed TN design. It is our hope that this Lecture Note will equip readers with a clear and unified understanding of the concept of TN rank, along with the necessary physical insight and intuition to support the selection, explainability, and deployment of tensor methods in both practical applications and educational contexts.
☆ Domain Borders Are There to Be Crossed With Federated Few-Shot Adaptation
Federated Learning has emerged as a leading paradigm for decentralized, privacy-preserving learning, particularly relevant in the era of interconnected edge devices equipped with sensors. However, the practical implementation of Federated Learning faces three primary challenges: the need for human involvement in costly data labelling processes for target adaptation, covariate shift in client device data collection due to environmental factors affecting sensors, leading to discrepancies between source and target samples, and the impracticality of continuous or regular model updates in resource-constrained environments due to limited data transmission capabilities and technical constraints on channel availability and energy efficiency. To tackle these issues, we expand upon an efficient and scalable Federated Learning framework tailored for real-world client adaptation in industrial settings. This framework leverages a pre-trained source model comprising a deep backbone, an adaptation module, and a classifier running on a powerful server. By freezing the backbone and classifier during client adaptation on resource-constrained devices, we allow the domain adaptive linear layer to handle target domain adaptation, thus minimizing overall computational overhead. Furthermore, this setup, designated as FedAcross+, is extended to encompass the processing of streaming data, thereby rendering the solution suitable for non-stationary environments. Extensive experimental results demonstrate the effectiveness of FedAcross+ in achieving competitive adaptation on low-end client devices with limited target samples, successfully addressing the challenge of domain shift. Moreover, our framework accommodates sporadic model updates within resource-constrained environments, ensuring practical and seamless deployment.
comment: Extension of http://dx.doi.org/10.5220/0012351900003654
☆ MTF-Grasp: A Multi-tier Federated Learning Approach for Robotic Grasping
Federated Learning (FL) is a promising machine learning paradigm that enables participating devices to train privacy-preserved and collaborative models. FL has proven its benefits for robotic manipulation tasks. However, grasping tasks lack exploration in such settings where robots train a global model without moving data and ensuring data privacy. The main challenge is that each robot learns from data that is nonindependent and identically distributed (non-IID) and of low quantity. This exhibits performance degradation, particularly in robotic grasping. Thus, in this work, we propose MTF-Grasp, a multi-tier FL approach for robotic grasping, acknowledging the unique challenges posed by the non-IID data distribution across robots, including quantitative skewness. MTF-Grasp harnesses data quality and quantity across robots to select a set of "top-level" robots with better data distribution and higher sample count. It then utilizes top-level robots to train initial seed models and distribute them to the remaining "low-level" robots, reducing the risk of model performance degradation in low-level robots. Our approach outperforms the conventional FL setup by up to 8% on the quantity-skewed Cornell and Jacquard grasping datasets.
comment: The work is accepted for presentation at IEEE SMC 2025
☆ Simulating Biases for Interpretable Fairness in Offline and Online Classifiers ECML
Predictive models often reinforce biases which were originally embedded in their training data, through skewed decisions. In such cases, mitigation methods are critical to ensure that, regardless of the prevailing disparities, model outcomes are adjusted to be fair. To assess this, datasets could be systematically generated with specific biases, to train machine learning classifiers. Then, predictive outcomes could aid in the understanding of this bias embedding process. Hence, an agent-based model (ABM), depicting a loan application process that represents various systemic biases across two demographic groups, was developed to produce synthetic datasets. Then, by applying classifiers trained on them to predict loan outcomes, we can assess how biased data leads to unfairness. This highlights a main contribution of this work: a framework for synthetic dataset generation with controllable bias injection. We also contribute with a novel explainability technique, which shows how mitigations affect the way classifiers leverage data features, via second-order Shapley values. In experiments, both offline and online learning approaches are employed. Mitigations are applied at different stages of the modelling pipeline, such as during pre-processing and in-processing.
comment: 17 pages, 2 figures, 1 equation, 3 tables: 1 in main body and 2 in the appendix. Submitted to the SynDAiTE: Synthetic Data for AI Trustworthiness and Evolution workshop from ECMLPKDD 2025, anonymized
☆ Deep Recurrence for Dynamical Segmentation Models
While biological vision systems rely heavily on feedback connections to iteratively refine perception, most artificial neural networks remain purely feedforward, processing input in a single static pass. In this work, we propose a predictive coding inspired feedback mechanism that introduces a recurrent loop from output to input, allowing the model to refine its internal state over time. We implement this mechanism within a standard U-Net architecture and introduce two biologically motivated operations, softmax projection and exponential decay, to ensure stability of the feedback loop. Through controlled experiments on a synthetic segmentation task, we show that the feedback model significantly outperforms its feedforward counterpart in noisy conditions and generalizes more effectively with limited supervision. Notably, feedback achieves above random performance with just two training examples, while the feedforward model requires at least four. Our findings demonstrate that feedback enhances robustness and data efficiency, and offer a path toward more adaptive and biologically inspired neural architectures. Code is available at: github.com/DCalhas/feedback_segmentation.
comment: 12 pages
☆ Adaptability in Multi-Agent Reinforcement Learning: A Framework and Unified Review
Multi-Agent Reinforcement Learning (MARL) has shown clear effectiveness in coordinating multiple agents across simulated benchmarks and constrained scenarios. However, its deployment in real-world multi-agent systems (MAS) remains limited, primarily due to the complex and dynamic nature of such environments. These challenges arise from multiple interacting sources of variability, including fluctuating agent populations, evolving task goals, and inconsistent execution conditions. Together, these factors demand that MARL algorithms remain effective under continuously changing system configurations and operational demands. To better capture and assess this capacity for adjustment, we introduce the concept of \textit{adaptability} as a unified and practically grounded lens through which to evaluate the reliability of MARL algorithms under shifting conditions, broadly referring to any changes in the environment dynamics that may occur during learning or execution. Centred on the notion of adaptability, we propose a structured framework comprising three key dimensions: learning adaptability, policy adaptability, and scenario-driven adaptability. By adopting this adaptability perspective, we aim to support more principled assessments of MARL performance beyond narrowly defined benchmarks. Ultimately, this survey contributes to the development of algorithms that are better suited for deployment in dynamic, real-world multi-agent systems.
☆ Large-Scale Graph Building in Dynamic Environments: Low Latency and High Quality
Learning and constructing large-scale graphs has attracted attention in recent decades, resulting in a rich literature that introduced various systems, tools, and algorithms. Grale is one of such tools that is designed for offline environments and is deployed in more than 50 different industrial settings at Google. Grale is widely applicable because of its ability to efficiently learn and construct a graph on datasets with multiple types of features. However, it is often the case that applications require the underlying data to evolve continuously and rapidly and the updated graph needs to be available with low latency. Such setting make the use of Grale prohibitive. While there are Approximate Nearest Neighbor (ANN) systems that handle dynamic updates with low latency, they are mostly limited to similarities over a single embedding. In this work, we introduce a system that inherits the advantages and the quality of Grale, and maintains a graph construction in a dynamic setting with tens of milliseconds of latency per request. We call the system Dynamic Grale Using ScaNN (Dynamic GUS). Our system has a wide range of applications with over 10 deployments at Google. One of the applications is in Android Security and Privacy, where Dynamic Grale Using ScaNN enables capturing harmful applications 4 times faster, before they can reach users.
☆ Wavelet-Enhanced Neural ODE and Graph Attention for Interpretable Energy Forecasting
Accurate forecasting of energy demand and supply is critical for optimizing sustainable energy systems, yet it is challenged by the variability of renewable sources and dynamic consumption patterns. This paper introduces a neural framework that integrates continuous-time Neural Ordinary Differential Equations (Neural ODEs), graph attention, multi-resolution wavelet transformations, and adaptive learning of frequencies to address the issues of time series prediction. The model employs a robust ODE solver, using the Runge-Kutta method, paired with graph-based attention and residual connections to better understand both structural and temporal patterns. Through wavelet-based feature extraction and adaptive frequency modulation, it adeptly captures and models diverse, multi-scale temporal dynamics. When evaluated across seven diverse datasets: ETTh1, ETTh2, ETTm1, ETTm2 (electricity transformer temperature), and Waste, Solar, and Hydro (renewable energy), this architecture consistently outperforms state-of-the-art baselines in various forecasting metrics, proving its robustness in capturing complex temporal dependencies. Furthermore, the model enhances interpretability through SHAP analysis, making it suitable for sustainable energy applications.
☆ A Variance-Reduced Cubic-Regularized Newton for Policy Optimization
In this paper, we study a second-order approach to policy optimization in reinforcement learning. Existing second-order methods often suffer from suboptimal sample complexity or rely on unrealistic assumptions about importance sampling. To overcome these limitations, we propose VR-CR-PN, a variance-reduced cubic-regularized policy Newton algorithm. To the best of our knowledge, this is the first algorithm that integrates Hessian-aided variance reduction with second-order policy optimization, effectively addressing the distribution shift problem and achieving best-known sample complexity under general nonconvex conditions but without the need for importance sampling. We theoretically establish that VR-CR-PN achieves a sample complexity of $\tilde{\mathcal{O}}(\epsilon^{-3})$ to reach an $\epsilon$-second-order stationary point, significantly improving upon the previous best result of $\tilde{\mathcal{O}}(\epsilon^{-3.5})$ under comparable assumptions. As an additional contribution, we introduce a novel Hessian estimator for the expected return function, which admits a uniform upper bound independent of the horizon length $H$, allowing the algorithm to achieve horizon-independent sample complexity.
comment: 13 pages, 1 figure
☆ Analysis of AI Techniques for Orchestrating Edge-Cloud Application Migration
Application migration in edge-cloud system enables high QoS and cost effective service delivery. However, automatically orchestrating such migration is typically solved with heuristic approaches. Starting from the Markov Decision Process (MDP), in this paper, we identify, analyze and compare selected state-of-the-art Artificial Intelligence (AI) planning and Reinforcement Learning (RL) approaches for solving the class of edge-cloud application migration problems that can be modeled as Towers of Hanoi (ToH) problems. We introduce a new classification based on state space definition and analyze the compared models also through this lense. The aim is to understand available techniques capable of orchestrating such application migration in emerging computing continuum environments.
☆ Towards High Supervised Learning Utility Training Data Generation: Data Pruning and Column Reordering KDD2025
Tabular data synthesis for supervised learning ('SL') model training is gaining popularity in industries such as healthcare, finance, and retail. Despite the progress made in tabular data generators, models trained with synthetic data often underperform compared to those trained with original data. This low SL utility of synthetic data stems from class imbalance exaggeration and SL data relationship overlooked by tabular generator. To address these challenges, we draw inspirations from techniques in emerging data-centric artificial intelligence and elucidate Pruning and ReOrdering ('PRRO'), a novel pipeline that integrates data-centric techniques into tabular data synthesis. PRRO incorporates data pruning to guide the table generator towards observations with high signal-to-noise ratio, ensuring that the class distribution of synthetic data closely matches that of the original data. Besides, PRRO employs a column reordering algorithm to align the data modeling structure of generators with that of SL models. These two modules enable PRRO to optimize SL utility of synthetic data. Empirical experiments on 22 public datasets show that synthetic data generated using PRRO enhances predictive performance compared to data generated without PRRO. Specifically, synthetic replacement of original data yields an average improvement of 26.74% and up to 871.46% improvement using PRRO, while synthetic appendant to original data results with PRRO-generated data results in an average improvement of 6.13% and up to 200.32%. Furthermore, experiments on six highly imbalanced datasets show that PRRO enables the generator to produce synthetic data with a class distribution that resembles the original data more closely, achieving a similarity improvement of 43%. Through PRRO, we foster a seamless integration of data synthesis to subsequent SL prediction, promoting quality and accessible data analysis.
comment: Accepted by Agentic & GenAI Evaluation KDD2025
☆ A Transfer Learning-Based Method for Water Body Segmentation in Remote Sensing Imagery: A Case Study of the Zhada Tulin Area
To address the prevalent challenges of domain shift and small sample sizes in remote sensing image water body segmentation, this study proposes and validates a two-stage transfer learning strategy based on the SegFormer model. The approach begins by training a foundational segmentation model on a diverse source domain, where it achieves an Intersection over Union (IoU) of 68.80% on its validation set, followed by fine-tuning on data from the distinct target domain. Focusing on the Zhada Tulin area in Tibet -- a region characterized by highly complex topography and spectral features -- the experimental results demonstrate that this strategy significantly boosts the IoU for the water body segmentation task from 25.50% (for direct transfer) to 64.84%. This not only effectively resolves the model performance degradation caused by domain discrepancy but also provides an effective technical paradigm for high-precision thematic information extraction in data-scarce and environmentally unique remote sensing scenarios.
comment: 13 pages, 6 figures, 2 tables
☆ Compression Method for Deep Diagonal State Space Model Based on $H^2$ Optimal Reduction
Deep learning models incorporating linear SSMs have gained attention for capturing long-range dependencies in sequential data. However, their large parameter sizes pose challenges for deployment on resource-constrained devices. In this study, we propose an efficient parameter reduction method for these models by applying $H^{2}$ model order reduction techniques from control theory to their linear SSM components. In experiments, the LRA benchmark results show that the model compression based on our proposed method outperforms an existing method using the Balanced Truncation, while successfully reducing the number of parameters in the SSMs to $1/32$ without sacrificing the performance of the original models.
comment: Accepted to IEEE Control Systems Letters
☆ ElasticMM: Efficient Multimodal LLMs Serving with Elastic Multimodal Parallelism
Multimodal large language models (MLLMs) extend LLMs to handle images, videos, and audio by incorporating feature extractors and projection modules. However, these additional components -- combined with complex inference pipelines and heterogeneous workloads -- introduce significant inference overhead. Therefore, efficiently serving MLLMs remains a major challenge. Current tightly coupled serving architectures struggle to distinguish between mixed request types or adapt parallelism strategies to different inference stages, leading to increased time-to-first-token (TTFT) latency and poor resource utilization. To address this, we propose Elastic Multimodal Parallelism (EMP), a new serving paradigm that elastically adapts to resource heterogeneity across request types and inference stages. Building upon EMP, we develop ElasticMM, an MLLM serving system that (1) separates requests into independent modality groups with dynamic resource allocation via a modality-aware load balancer; (2) decouples inference stages and enables parallelism adjustment and adaptive scaling via elastic partition scheduling; and (3) improves inference efficiency through unified multimodal prefix caching and non-blocking encoding. Experiments on diverse real-world datasets show that ElasticMM outperforms state-of-the-art (SOTA) serving systems, reducing TTFT by up to 4.2x and achieving 3.2-4.5x higher throughput while meeting service-level objectives (SLOs).
☆ PRISM: Fine-Grained Paper-to-Paper Retrieval with Multi-Aspect-Aware Query Optimization
Scientific paper retrieval, particularly framed as document-to-document retrieval, aims to identify relevant papers in response to a long-form query paper, rather than a short query string. Previous approaches to this task have focused on abstracts, embedding them into dense vectors as surrogates for full documents and calculating similarity across them, although abstracts provide only sparse and high-level summaries. To address this, we propose PRISM, a novel document-to-document retrieval method that introduces multiple, fine-grained representations for both the query and candidate papers. In particular, each query paper is decomposed into multiple aspect-specific views and individually embedded, which are then matched against candidate papers similarity segmented to consider their multifaceted dimensions. Moreover, we present SciFullBench, a novel benchmark in which the complete and segmented context of full papers for both queries and candidates is available. Then, experimental results show that PRISM improves performance by an average of 4.3% over existing retrieval baselines.
☆ Lightweight Model for Poultry Disease Detection from Fecal Images Using Multi-Color Space Feature Optimization and Machine Learning
Poultry farming is a vital component of the global food supply chain, yet it remains highly vulnerable to infectious diseases such as coccidiosis, salmonellosis, and Newcastle disease. This study proposes a lightweight machine learning-based approach to detect these diseases by analyzing poultry fecal images. We utilize multi-color space feature extraction (RGB, HSV, LAB) and explore a wide range of color, texture, and shape-based descriptors, including color histograms, local binary patterns (LBP), wavelet transforms, and edge detectors. Through a systematic ablation study and dimensionality reduction using PCA and XGBoost feature selection, we identify a compact global feature set that balances accuracy and computational efficiency. An artificial neural network (ANN) classifier trained on these features achieved 95.85% accuracy while requiring no GPU and only 638 seconds of execution time in Google Colab. Compared to deep learning models such as Xception and MobileNetV3, our proposed model offers comparable accuracy with drastically lower resource usage. This work demonstrates a cost-effective, interpretable, and scalable alternative to deep learning for real-time poultry disease detection in low-resource agricultural settings.
☆ On the Efficiency of Training Robust Decision Trees
As machine learning gets adopted into the industry quickly, trustworthiness is increasingly in focus. Yet, efficiency and sustainability of robust training pipelines still have to be established. In this work, we consider a simple pipeline for training adversarially robust decision trees and investigate the efficiency of each step. Our pipeline consists of three stages. Firstly, we choose the perturbation size automatically for each dataset. For that, we introduce a simple algorithm, instead of relying on intuition or prior work. Moreover, we show that the perturbation size can be estimated from smaller models than the one intended for full training, and thus significant gains in efficiency can be achieved. Secondly, we train state-of-the-art adversarial training methods and evaluate them regarding both their training time and adversarial accuracy. Thirdly, we certify the robustness of each of the models thus obtained and investigate the time required for this. We find that verification time, which is critical to the efficiency of the full pipeline, is not correlated with training time.
comment: Presented as a poster at SAIV 2025
☆ Towards Applying Large Language Models to Complement Single-Cell Foundation Models
Single-cell foundation models such as scGPT represent a significant advancement in single-cell omics, with an ability to achieve state-of-the-art performance on various downstream biological tasks. However, these models are inherently limited in that a vast amount of information in biology exists as text, which they are unable to leverage. There have therefore been several recent works that propose the use of LLMs as an alternative to single-cell foundation models, achieving competitive results. However, there is little understanding of what factors drive this performance, along with a strong focus on using LLMs as an alternative, rather than complementary approach to single-cell foundation models. In this study, we therefore investigate what biological insights contribute toward the performance of LLMs when applied to single-cell data, and introduce scMPT; a model which leverages synergies between scGPT, and single-cell representations from LLMs that capture these insights. scMPT demonstrates stronger, more consistent performance than either of its component models, which frequently have large performance gaps between each other across datasets. We also experiment with alternate fusion methods, demonstrating the potential of combining specialized reasoning models with scGPT to improve performance. This study ultimately showcases the potential for LLMs to complement single-cell foundation models and drive improvements in single-cell analysis.
☆ Memory-Efficient Personalization of Text-to-Image Diffusion Models via Selective Optimization Strategies
Memory-efficient personalization is critical for adapting text-to-image diffusion models while preserving user privacy and operating within the limited computational resources of edge devices. To this end, we propose a selective optimization framework that adaptively chooses between backpropagation on low-resolution images (BP-low) and zeroth-order optimization on high-resolution images (ZO-high), guided by the characteristics of the diffusion process. As observed in our experiments, BP-low efficiently adapts the model to target-specific features, but suffers from structural distortions due to resolution mismatch. Conversely, ZO-high refines high-resolution details with minimal memory overhead but faces slow convergence when applied without prior adaptation. By complementing both methods, our framework leverages BP-low for effective personalization while using ZO-high to maintain structural consistency, achieving memory-efficient and high-quality fine-tuning. To maximize the efficacy of both BP-low and ZO-high, we introduce a timestep-aware probabilistic function that dynamically selects the appropriate optimization strategy based on diffusion timesteps. This function mitigates the overfitting from BP-low at high timesteps, where structural information is critical, while ensuring ZO-high is applied more effectively as training progresses. Experimental results demonstrate that our method achieves competitive performance while significantly reducing memory consumption, enabling scalable, high-quality on-device personalization without increasing inference latency.
☆ (Almost) Free Modality Stitching of Foundation Models
Foundation multi-modal models are often designed by stitching of multiple existing pretrained uni-modal models: for example, an image classifier with an autoregressive text model. This stitching process is performed by training a connector module that aims to align the representation-representation or representation-input spaces of these uni-modal models. However, given the complexity of training such connectors on large scale web-based datasets coupled with the ever-increasing number of available pretrained uni-modal models, the task of uni-modal models selection and subsequent connector module training becomes computationally demanding. To address this under-studied critical problem, we propose Hypernetwork Model Alignment (Hyma), a novel all-in-one solution for optimal uni-modal model selection and connector training by leveraging hypernetworks. Specifically, our framework utilizes the parameter prediction capability of a hypernetwork to obtain jointly trained connector modules for $N \times M$ combinations of uni-modal models. In our experiments, Hyma reduces the optimal uni-modal model pair search cost by $10\times$ (averaged across all experiments), while matching the ranking and trained connector performance obtained via grid search across a suite of diverse multi-modal benchmarks.
comment: Pre-print
☆ Forecasting Coccidioidomycosis (Valley Fever) in Arizona: A Graph Neural Network Approach
Coccidioidomycosis, commonly known as Valley Fever, remains a significant public health concern in endemic regions of the southwestern United States. This study develops the first graph neural network (GNN) model for forecasting Valley Fever incidence in Arizona. The model integrates surveillance case data with environmental predictors using graph structures, including soil conditions, atmospheric variables, agricultural indicators, and air quality metrics. Our approach explores correlation-based relationships among variables influencing disease transmission. The model captures critical delays in disease progression through lagged effects, enhancing its capacity to reflect complex temporal dependencies in disease ecology. Results demonstrate that the GNN architecture effectively models Valley Fever trends and provides insights into key environmental drivers of disease incidence. These findings can inform early warning systems and guide resource allocation for disease prevention efforts in high-risk areas.
☆ Effects of structural properties of neural networks on machine learning performance
In recent years, graph-based machine learning techniques, such as reinforcement learning and graph neural networks, have garnered significant attention. While some recent studies have started to explore the relationship between the graph structure of neural networks and their predictive performance, they often limit themselves to a narrow range of model networks, particularly lacking mesoscale structures such as communities. Our work advances this area by conducting a more comprehensive investigation, incorporating realistic network structures characterized by heterogeneous degree distributions and community structures, which are typical characteristics of many real networks. These community structures offer a nuanced perspective on network architecture. Our analysis employs model networks such as random and scale-free networks, alongside a comparison with a biological neural network and its subsets for more detailed analysis. We examine the impact of these structural attributes on the performance of image classification tasks. Our findings reveal that structural properties do affect performance to some extent. Specifically, networks featuring coherent, densely interconnected communities demonstrate enhanced learning capabilities. The comparison with the biological neural network emphasizes the relevance of our findings to real-world structures, suggesting an intriguing connection worth further exploration. This study contributes meaningfully to network science and machine learning, providing insights that could inspire the design of more biologically informed neural networks.
comment: 9 pages, 6 figures
☆ Compliance Minimization via Physics-Informed Gaussian Processes
Machine learning (ML) techniques have recently gained significant attention for solving compliance minimization (CM) problems. However, these methods typically provide poor feature boundaries, are very expensive, and lack a systematic mechanism to control the design complexity. Herein, we address these limitations by proposing a mesh-free and simultaneous framework based on physics-informed Gaussian processes (GPs). In our approach, we parameterize the design and state variables with GP priors which have independent kernels but share a multi-output neural network (NN) as their mean function. The architecture of this NN is based on Parametric Grid Convolutional Attention Networks (PGCANs) which not only mitigate spectral bias issues, but also provide an interpretable mechanism to control design complexity. We estimate all the parameters of our GP-based representations by simultaneously minimizing the compliance, total potential energy, and residual of volume fraction constraint. Importantly, our loss function exclude all data-based residuals as GPs automatically satisfy them. We also develop computational schemes based on curriculum training and numerical integration to increase the efficiency and robustness of our approach which is shown to (1) produce super-resolution topologies with fast convergence, (2) achieve smaller compliance and less gray area fraction compared to traditional numerical methods, (3) provide control over fine-scale features, and (4) outperform competing ML-based methods.
☆ A Brain Tumor Segmentation Method Based on CLIP and 3D U-Net with Cross-Modal Semantic Guidance and Multi-Level Feature Fusion
Precise segmentation of brain tumors from magnetic resonance imaging (MRI) is essential for neuro-oncology diagnosis and treatment planning. Despite advances in deep learning methods, automatic segmentation remains challenging due to tumor morphological heterogeneity and complex three-dimensional spatial relationships. Current techniques primarily rely on visual features extracted from MRI sequences while underutilizing semantic knowledge embedded in medical reports. This research presents a multi-level fusion architecture that integrates pixel-level, feature-level, and semantic-level information, facilitating comprehensive processing from low-level data to high-level concepts. The semantic-level fusion pathway combines the semantic understanding capabilities of Contrastive Language-Image Pre-training (CLIP) models with the spatial feature extraction advantages of 3D U-Net through three mechanisms: 3D-2D semantic bridging, cross-modal semantic guidance, and semantic-based attention mechanisms. Experimental validation on the BraTS 2020 dataset demonstrates that the proposed model achieves an overall Dice coefficient of 0.8567, representing a 4.8% improvement compared to traditional 3D U-Net, with a 7.3% Dice coefficient increase in the clinically important enhancing tumor (ET) region.
comment: 13 pages,6 figures
☆ Text-Driven Causal Representation Learning for Source-Free Domain Generalization
Deep learning often struggles when training and test data distributions differ. Traditional domain generalization (DG) tackles this by including data from multiple source domains, which is impractical due to expensive data collection and annotation. Recent vision-language models like CLIP enable source-free domain generalization (SFDG) by using text prompts to simulate visual representations, reducing data demands. However, existing SFDG methods struggle with domain-specific confounders, limiting their generalization capabilities. To address this issue, we propose TDCRL (\textbf{T}ext-\textbf{D}riven \textbf{C}ausal \textbf{R}epresentation \textbf{L}earning), the first method to integrate causal inference into the SFDG setting. TDCRL operates in two steps: first, it employs data augmentation to generate style word vectors, combining them with class information to generate text embeddings to simulate visual representations; second, it trains a causal intervention network with a confounder dictionary to extract domain-invariant features. Grounded in causal learning, our approach offers a clear and effective mechanism to achieve robust, domain-invariant features, ensuring robust generalization. Extensive experiments on PACS, VLCS, OfficeHome, and DomainNet show state-of-the-art performance, proving TDCRL effectiveness in SFDG.
comment: Under Review
☆ Rethinking Inductive Bias in Geographically Neural Network Weighted Regression
Inductive bias is a key factor in spatial regression models, determining how well a model can learn from limited data and capture spatial patterns. This work revisits the inductive biases in Geographically Neural Network Weighted Regression (GNNWR) and identifies limitations in current approaches for modeling spatial non-stationarity. While GNNWR extends traditional Geographically Weighted Regression by using neural networks to learn spatial weighting functions, existing implementations are often restricted by fixed distance-based schemes and limited inductive bias. We propose to generalize GNNWR by incorporating concepts from convolutional neural networks, recurrent neural networks, and transformers, introducing local receptive fields, sequential context, and self-attention into spatial regression. Through extensive benchmarking on synthetic spatial datasets with varying heterogeneity, noise, and sample sizes, we show that GNNWR outperforms classic methods in capturing nonlinear and complex spatial relationships. Our results also reveal that model performance depends strongly on data characteristics, with local models excelling in highly heterogeneous or small-sample scenarios, and global models performing better with larger, more homogeneous data. These findings highlight the importance of inductive bias in spatial modeling and suggest future directions, including learnable spatial weighting functions, hybrid neural architectures, and improved interpretability for models handling non-stationary spatial data.
☆ Radial Neighborhood Smoothing Recommender System NeurIPS 2025
Recommender systems inherently exhibit a low-rank structure in latent space. A key challenge is to define meaningful and measurable distances in the latent space to capture user-user, item-item, user-item relationships effectively. In this work, we establish that distances in the latent space can be systematically approximated using row-wise and column-wise distances in the observed matrix, providing a novel perspective on distance estimation. To refine the distance estimation, we introduce the correction based on empirical variance estimator to account for noise-induced non-centrality. The novel distance estimation enables a more structured approach to constructing neighborhoods, leading to the Radial Neighborhood Estimator (RNE), which constructs neighborhoods by including both overlapped and partially overlapped user-item pairs and employs neighborhood smoothing via localized kernel regression to improve imputation accuracy. We provide the theoretical asymptotic analysis for the proposed estimator. We perform evaluations on both simulated and real-world datasets, demonstrating that RNE achieves superior performance compared to existing collaborative filtering and matrix factorization methods. While our primary focus is on distance estimation in latent space, we find that RNE also mitigates the ``cold-start'' problem.
comment: 34 pages, 2 figures. Submitted to NeurIPS 2025
☆ Hierarchical Job Classification with Similarity Graph Integration
In the dynamic realm of online recruitment, accurate job classification is paramount for optimizing job recommendation systems, search rankings, and labor market analyses. As job markets evolve, the increasing complexity of job titles and descriptions necessitates sophisticated models that can effectively leverage intricate relationships within job data. Traditional text classification methods often fall short, particularly due to their inability to fully utilize the hierarchical nature of industry categories. To address these limitations, we propose a novel representation learning and classification model that embeds jobs and hierarchical industry categories into a latent embedding space. Our model integrates the Standard Occupational Classification (SOC) system and an in-house hierarchical taxonomy, Carotene, to capture both graph and hierarchical relationships, thereby improving classification accuracy. By embedding hierarchical industry categories into a shared latent space, we tackle cold start issues and enhance the dynamic matching of candidates to job opportunities. Extensive experimentation on a large-scale dataset of job postings demonstrates the model's superior ability to leverage hierarchical structures and rich semantic features, significantly outperforming existing methods. This research provides a robust framework for improving job classification accuracy, supporting more informed decision-making in the recruitment industry.
☆ Iceberg: Enhancing HLS Modeling with Synthetic Data
Deep learning-based prediction models for High-Level Synthesis (HLS) of hardware designs often struggle to generalize. In this paper, we study how to close the generalizability gap of these models through pretraining on synthetic data and introduce Iceberg, a synthetic data augmentation approach that expands both large language model (LLM)-generated programs and weak labels of unseen design configurations. Our weak label generation method is integrated with an in-context model architecture, enabling meta-learning from actual and proximate labels. Iceberg improves the geometric mean modeling accuracy by $86.4\%$ when adapt to six real-world applications with few-shot examples and achieves a $2.47\times$ and a $1.12\times$ better offline DSE performance when adapting to two different test datasets. Our open-sourced code is here: \href{https://github.com/UCLA-VAST/iceberg}{https://github.com/UCLA-VAST/iceberg}
comment: 9 pages. accepted to ICLAD'25
☆ Long-Tailed Data Classification by Increasing and Decreasing Neurons During Training
In conventional deep learning, the number of neurons typically remains fixed during training. However, insights from biology suggest that the human hippocampus undergoes continuous neuron generation and pruning of neurons over the course of learning, implying that a flexible allocation of capacity can contribute to enhance performance. Real-world datasets often exhibit class imbalance situations where certain classes have far fewer samples than others, leading to significantly reduce recognition accuracy for minority classes when relying on fixed size networks.To address the challenge, we propose a method that periodically adds and removes neurons during training, thereby boosting representational power for minority classes. By retaining critical features learned from majority classes while selectively increasing neurons for underrepresented classes, our approach dynamically adjusts capacity during training. Importantly, while the number of neurons changes throughout training, the final network size and structure remain unchanged, ensuring efficiency and compatibility with deployment.Furthermore, by experiments on three different datasets and five representative models, we demonstrate that the proposed method outperforms fixed size networks and shows even greater accuracy when combined with other imbalance-handling techniques. Our results underscore the effectiveness of dynamic, biologically inspired network designs in improving performance on class-imbalanced data.
☆ Memorization Sinks: Isolating Memorization during LLM Training
Large language models are susceptible to memorizing repeated sequences, posing privacy and copyright concerns. A popular mitigation strategy is to remove memorized information from specific neurons post-hoc. However, such approaches have shown limited success so far. In a controlled setting, we show that the memorization of natural sequences (those that resemble linguistically plausible text) become mechanistically entangled with general language abilities, thereby becoming challenging to remove post-hoc. In this work, we put forward a new paradigm of MemSinks that promotes isolation of memorization by design. We leverage a sequence identifier that activates a unique set of memorization neurons for each sequence across repetitions. By analyzing the dynamics of learning and forgetting, we argue that MemSinks facilitates isolation of memorized content, making it easier to remove without compromising general language capabilities. We implement MemSinks at the billion-parameter and billion-token scale, and observe both effective isolation and strong generalization. To our knowledge, this is the first proof-of-concept on real data demonstrating that simultaneous generalization and isolation is achievable. We open-source our code at http://github.com/grghosal/MemSinks.
comment: Accepted at the 2025 International Conference of Machine Learning
☆ Mechanistic Interpretability of LoRA-Adapted Language Models for Nuclear Reactor Safety Applications
The integration of Large Language Models (LLMs) into safety-critical domains, such as nuclear engineering, necessitates a deep understanding of their internal reasoning processes. This paper presents a novel methodology for interpreting how an LLM encodes and utilizes domain-specific knowledge, using a Boiling Water Reactor system as a case study. We adapted a general-purpose LLM (Gemma-3-1b-it) to the nuclear domain using a parameter-efficient fine-tuning technique known as Low-Rank Adaptation. By comparing the neuron activation patterns of the base model to those of the fine-tuned model, we identified a sparse set of neurons whose behavior was significantly altered during the adaptation process. To probe the causal role of these specialized neurons, we employed a neuron silencing technique. Our results demonstrate that while silencing most of these specialized neurons individually did not produce a statistically significant effect, deactivating the entire group collectively led to a statistically significant degradation in task performance. Qualitative analysis further revealed that silencing these neurons impaired the model's ability to generate detailed, contextually accurate technical information. This paper provides a concrete methodology for enhancing the transparency of an opaque black-box model, allowing domain expertise to be traced to verifiable neural circuits. This offers a pathway towards achieving nuclear-grade artificial intelligence (AI) assurance, addressing the verification and validation challenges mandated by nuclear regulatory frameworks (e.g., 10 CFR 50 Appendix B), which have limited AI deployment in safety-critical nuclear operations.
comment: Submitted to Nuclear Technology. 22 pages, 2 tables, 4 figures
☆ Aligning Generative Speech Enhancement with Human Preferences via Direct Preference Optimization
This work investigates speech enhancement (SE) from the perspective of language models (LMs). We propose a novel method that leverages Direct Preference Optimization (DPO) to improve the perceptual quality of enhanced speech. Using UTMOS, a neural MOS prediction model, as a proxy for human ratings, our approach guides optimization toward perceptually preferred outputs. This differs from existing LM-based SE methods that focus on maximizing the likelihood of clean speech tokens, which may misalign with human perception and degrade quality despite low prediction error. Experiments on the 2020 Deep Noise Suppression Challenge test sets demonstrate that applying DPO to a pretrained LM-based SE model yields consistent improvements across various speech quality metrics, with relative gains of up to 56%. To our knowledge, this is the first application of DPO to SE and the first to incorporate proxy perceptual feedback into LM-based SE training, pointing to a promising direction for perceptually aligned SE.
☆ Extracting Cause-Effect Pairs from a Sentence with a Dependency-Aware Transformer Model
Extracting cause and effect phrases from a sentence is an important NLP task, with numerous applications in various domains, including legal, medical, education, and scientific research. There are many unsupervised and supervised methods proposed for solving this task. Among these, unsupervised methods utilize various linguistic tools, including syntactic patterns, dependency tree, dependency relations, etc. among different sentential units for extracting the cause and effect phrases. On the other hand, the contemporary supervised methods use various deep learning based mask language models equipped with a token classification layer for extracting cause and effect phrases. Linguistic tools, specifically, dependency tree, which organizes a sentence into different semantic units have been shown to be very effective for extracting semantic pairs from a sentence, but existing supervised methods do not have any provision for utilizing such tools within their model framework. In this work, we propose DepBERT, which extends a transformer-based model by incorporating dependency tree of a sentence within the model framework. Extensive experiments over three datasets show that DepBERT is better than various state-of-the art supervised causality extraction methods.
☆ MixLoRA-DSI: Dynamically Expandable Mixture-of-LoRA Experts for Rehearsal-Free Generative Retrieval over Dynamic Corpora
Continually updating model-based indexes in generative retrieval with new documents remains challenging, as full retraining is computationally expensive and impractical under resource constraints. We propose MixLoRA-DSI, a novel framework that combines an expandable mixture of Low-Rank Adaptation experts with a layer-wise out-of-distribution (OOD)-driven expansion strategy. Instead of allocating new experts for each new corpus, our proposed expansion strategy enables sublinear parameter growth by selectively introducing new experts only when significant number of OOD documents are detected. Experiments on NQ320k and MS MARCO Passage demonstrate that MixLoRA-DSI outperforms full-model update baselines, with minimal parameter overhead and substantially lower training costs.
☆ Advanced U-Net Architectures with CNN Backbones for Automated Lung Cancer Detection and Segmentation in Chest CT Images
This study investigates the effectiveness of U-Net architectures integrated with various convolutional neural network (CNN) backbones for automated lung cancer detection and segmentation in chest CT images, addressing the critical need for accurate diagnostic tools in clinical settings. A balanced dataset of 832 chest CT images (416 cancerous and 416 non-cancerous) was preprocessed using Contrast Limited Adaptive Histogram Equalization (CLAHE) and resized to 128x128 pixels. U-Net models were developed with three CNN backbones: ResNet50, VGG16, and Xception, to segment lung regions. After segmentation, CNN-based classifiers and hybrid models combining CNN feature extraction with traditional machine learning classifiers (Support Vector Machine, Random Forest, and Gradient Boosting) were evaluated using 5-fold cross-validation. Metrics included accuracy, precision, recall, F1-score, Dice coefficient, and ROC-AUC. U-Net with ResNet50 achieved the best performance for cancerous lungs (Dice: 0.9495, Accuracy: 0.9735), while U-Net with VGG16 performed best for non-cancerous segmentation (Dice: 0.9532, Accuracy: 0.9513). For classification, the CNN model using U-Net with Xception achieved 99.1 percent accuracy, 99.74 percent recall, and 99.42 percent F1-score. The hybrid CNN-SVM-Xception model achieved 96.7 percent accuracy and 97.88 percent F1-score. Compared to prior methods, our framework consistently outperformed existing models. In conclusion, combining U-Net with advanced CNN backbones provides a powerful method for both segmentation and classification of lung cancer in CT scans, supporting early diagnosis and clinical decision-making.
comment: This manuscript has 20 pages and 10 figures. It is submitted to the Journal 'Scientific Reports'
☆ Algorithm Development in Neural Networks: Insights from the Streaming Parity Task
Even when massively overparameterized, deep neural networks show a remarkable ability to generalize. Research on this phenomenon has focused on generalization within distribution, via smooth interpolation. Yet in some settings neural networks also learn to extrapolate to data far beyond the bounds of the original training set, sometimes even allowing for infinite generalization, implying that an algorithm capable of solving the task has been learned. Here we undertake a case study of the learning dynamics of recurrent neural networks (RNNs) trained on the streaming parity task in order to develop an effective theory of algorithm development. The streaming parity task is a simple but nonlinear task defined on sequences up to arbitrary length. We show that, with sufficient finite training experience, RNNs exhibit a phase transition to perfect infinite generalization. Using an effective theory for the representational dynamics, we find an implicit representational merger effect which can be interpreted as the construction of a finite automaton that reproduces the task. Overall, our results disclose one mechanism by which neural networks can generalize infinitely from finite training experience.
comment: 28 pages, 20 figures
♻ ☆ Expert-level validation of AI-generated medical text with scalable language models
With the growing use of language models (LMs) in clinical environments, there is an immediate need to evaluate the accuracy and safety of LM-generated medical text. Currently, such evaluation relies solely on manual physician review. However, detecting errors in LM-generated text is challenging because 1) manual review is costly and 2) expert-composed reference outputs are often unavailable in real-world settings. While the "LM-as-judge" paradigm (a LM evaluating another LM) offers scalable evaluation, even frontier LMs can miss subtle but clinically significant errors. To address these challenges, we propose MedVAL, a self-supervised framework that leverages synthetic data to train evaluator LMs to assess whether LM-generated medical outputs are factually consistent with inputs, without requiring physician labels or reference outputs. To evaluate LM performance, we introduce MedVAL-Bench, a dataset containing 840 outputs annotated by physicians, following a physician-defined taxonomy of risk levels and error categories. Across 6 diverse medical tasks and 10 state-of-the-art LMs spanning open-source, proprietary, and medically adapted models, MedVAL fine-tuning significantly improves (p < 0.001) alignment with physicians on both seen and unseen tasks, increasing average F1 scores from 66% to 83%, with per-sample safety classification scores up to 86%. MedVAL improves the performance of even the best-performing proprietary LM (GPT-4o) by 8%. To support a scalable, risk-aware pathway towards clinical integration, we open-source the 1) codebase (https://github.com/StanfordMIMI/MedVAL), 2) MedVAL-Bench (https://huggingface.co/datasets/stanfordmimi/MedVAL-Bench), and 3) MedVAL-4B (https://huggingface.co/stanfordmimi/MedVAL-4B), the best-performing open-source LM. Our research provides the first evidence of LMs approaching expert-level validation ability for medical text.
♻ ☆ Ark: An Open-source Python-based Framework for Robot Learning
Robotics has made remarkable hardware strides-from DARPA's Urban and Robotics Challenges to the first humanoid-robot kickboxing tournament-yet commercial autonomy still lags behind progress in machine learning. A major bottleneck is software: current robot stacks demand steep learning curves, low-level C/C++ expertise, fragmented tooling, and intricate hardware integration, in stark contrast to the Python-centric, well-documented ecosystems that propelled modern AI. We introduce ARK, an open-source, Python-first robotics framework designed to close that gap. ARK presents a Gym-style environment interface that allows users to collect data, preprocess it, and train policies using state-of-the-art imitation-learning algorithms (e.g., ACT, Diffusion Policy) while seamlessly toggling between high-fidelity simulation and physical robots. A lightweight client-server architecture provides networked publisher-subscriber communication, and optional C/C++ bindings ensure real-time performance when needed. ARK ships with reusable modules for control, SLAM, motion planning, system identification, and visualization, along with native ROS interoperability. Comprehensive documentation and case studies-from manipulation to mobile navigation-demonstrate rapid prototyping, effortless hardware swapping, and end-to-end pipelines that rival the convenience of mainstream machine-learning workflows. By unifying robotics and AI practices under a common Python umbrella, ARK lowers entry barriers and accelerates research and commercial deployment of autonomous robots.
♻ ☆ Visual Test-time Scaling for GUI Agent Grounding ICCV2025
We introduce RegionFocus, a visual test-time scaling approach for Vision Language Model Agents. Understanding webpages is challenging due to the visual complexity of GUI images and the large number of interface elements, making accurate action selection difficult. Our approach dynamically zooms in on relevant regions, reducing background clutter and improving grounding accuracy. To support this process, we propose an image-as-map mechanism that visualizes key landmarks at each step, providing a transparent action record and enables the agent to effectively choose among action candidates. Even with a simple region selection strategy, we observe significant performance gains of 28+\% on Screenspot-pro and 24+\% on WebVoyager benchmarks on top of two state-of-the-art open vision language model agents, UI-TARS and Qwen2.5-VL, highlighting the effectiveness of visual test-time scaling in interactive settings. We achieve a new state-of-the-art grounding performance of 61.6\% on the ScreenSpot-Pro benchmark by applying RegionFocus to a Qwen2.5-VL-72B model. Our code will be released publicly at https://github.com/tiangeluo/RegionFocus.
comment: ICCV2025, https://github.com/tiangeluo/RegionFocus
♻ ☆ A Unified View on Learning Unnormalized Distributions via Noise-Contrastive Estimation ICML 2025
This paper studies a family of estimators based on noise-contrastive estimation (NCE) for learning unnormalized distributions. The main contribution of this work is to provide a unified perspective on various methods for learning unnormalized distributions, which have been independently proposed and studied in separate research communities, through the lens of NCE. This unified view offers new insights into existing estimators. Specifically, for exponential families, we establish the finite-sample convergence rates of the proposed estimators under a set of regularity assumptions, most of which are new.
comment: 31 pages, 2 figures. ICML 2025
♻ ☆ Improved Offline Contextual Bandits with Second-Order Bounds: Betting and Freezing COLT 2025
We consider off-policy selection and learning in contextual bandits, where the learner aims to select or train a reward-maximizing policy using data collected by a fixed behavior policy. Our contribution is two-fold. First, we propose a novel off-policy selection method that leverages a new betting-based confidence bound applied to an inverse propensity weight sequence. Our theoretical analysis reveals that this method achieves a significantly improved, variance-adaptive guarantee over prior work. Second, we propose a novel and generic condition on the optimization objective for off-policy learning that strikes a different balance between bias and variance. One special case, which we call freezing, tends to induce low variance, which is preferred in small-data regimes. Our analysis shows that it matches the best existing guarantees. In our empirical study, our selection method outperforms existing methods, and freezing exhibits improved performance in small-sample regimes.
comment: 39 pages, 10 figures. COLT 2025
♻ ☆ On the Robustness Tradeoff in Fine-Tuning ICCV 2025
Fine-tuning has become the standard practice for adapting pre-trained models to downstream tasks. However, the impact on model robustness is not well understood. In this work, we characterize the robustness-accuracy trade-off in fine-tuning. We evaluate the robustness and accuracy of fine-tuned models over 6 benchmark datasets and 7 different fine-tuning strategies. We observe a consistent trade-off between adversarial robustness and accuracy. Peripheral updates such as BitFit are more effective for simple tasks -- over 75% above the average measured by the area under the Pareto frontiers on CIFAR-10 and CIFAR-100. In contrast, fine-tuning information-heavy layers, such as attention layers via Compacter, achieves a better Pareto frontier on more complex tasks -- 57.5% and 34.6% above the average on Caltech-256 and CUB-200, respectively. Lastly, we observe that the robustness of fine-tuning against out-of-distribution data closely tracks accuracy. These insights emphasize the need for robustness-aware fine-tuning to ensure reliable real-world deployments.
comment: Accepted to International Conference on Computer Vision, ICCV 2025
♻ ☆ Enabling Advanced Land Cover Analytics: An Integrated Data Extraction Pipeline for Predictive Modeling with the Dynamic World Dataset
Understanding land cover holds considerable potential for a myriad of practical applications, particularly as data accessibility transitions from being exclusive to governmental and commercial entities to now including the broader research community. Nevertheless, although the data is accessible to any community member interested in exploration, there exists a formidable learning curve and no standardized process for accessing, pre-processing, and leveraging the data for subsequent tasks. In this study, we democratize this data by presenting a flexible and efficient end to end pipeline for working with the Dynamic World dataset, a cutting-edge near-real-time land use/land cover (LULC) dataset. This includes a pre-processing and representation framework which tackles noise removal, efficient extraction of large amounts of data, and re-representation of LULC data in a format well suited for several downstream tasks. To demonstrate the power of our pipeline, we use it to extract data for an urbanization prediction problem and build a suite of machine learning models with excellent performance. This task is easily generalizable to the prediction of any type of land cover and our pipeline is also compatible with a series of other downstream tasks.
♻ ☆ Random Erasing vs. Model Inversion: A Promising Defense or a False Hope?
Model Inversion (MI) attacks pose a significant privacy threat by reconstructing private training data from machine learning models. While existing defenses primarily concentrate on model-centric approaches, the impact of data on MI robustness remains largely unexplored. In this work, we explore Random Erasing (RE), a technique traditionally used for improving model generalization under occlusion, and uncover its surprising effectiveness as a defense against MI attacks. Specifically, our novel feature space analysis shows that models trained with RE-images introduce a significant discrepancy between the features of MI-reconstructed images and those of the private data. At the same time, features of private images remain distinct from other classes and well-separated from different classification regions. These effects collectively degrade MI reconstruction quality and attack accuracy while maintaining reasonable natural accuracy. Furthermore, we explore two critical properties of RE including Partial Erasure and Random Location. Partial Erasure prevents the model from observing entire objects during training. We find this has a significant impact on MI, which aims to reconstruct the entire objects. Random Location of erasure plays a crucial role in achieving a strong privacy-utility trade-off. Our findings highlight RE as a simple yet effective defense mechanism that can be easily integrated with existing privacy-preserving techniques. Extensive experiments across 37 setups demonstrate that our method achieves state-of-the-art (SOTA) performance in the privacy-utility trade-off. The results consistently demonstrate the superiority of our defense over existing methods across different MI attacks, network architectures, and attack configurations. For the first time, we achieve a significant degradation in attack accuracy without a decrease in utility for some configurations.
comment: Accepted in Transactions on Machine Learning Research (TMLR). First two authors contributed equally
♻ ☆ An Interoperable Machine Learning Pipeline for Pediatric Obesity Risk Estimation ML4H
Reliable prediction of pediatric obesity can offer a valuable resource to providers, helping them engage in timely preventive interventions before the disease is established. Many efforts have been made to develop ML-based predictive models of obesity, and some studies have reported high predictive performances. However, no commonly used clinical decision support tool based on existing ML models currently exists. This study presents a novel end-to-end pipeline specifically designed for pediatric obesity prediction, which supports the entire process of data extraction, inference, and communication via an API or a user interface. While focusing only on routinely recorded data in pediatric electronic health records (EHRs), our pipeline uses a diverse expert-curated list of medical concepts to predict the 1-3 years risk of developing obesity. Furthermore, by using the Fast Healthcare Interoperability Resources (FHIR) standard in our design procedure, we specifically target facilitating low-effort integration of our pipeline with different EHR systems. In our experiments, we report the effectiveness of the predictive model as well as its alignment with the feedback from various stakeholders, including ML scientists, providers, health IT personnel, health administration representatives, and patient group representatives.
comment: This paper has been accepted in Machine Learning for Health (ML4H) Symposium. Link: https://proceedings.mlr.press/v259/fayyaz25a.html
♻ ☆ Discrimination-free Insurance Pricing with Privatized Sensitive Attributes
Fairness has emerged as a critical consideration in the landscape of machine learning algorithms, particularly as AI continues to transform decision-making across societal domains. To ensure that these algorithms are free from bias and do not discriminate against individuals based on sensitive attributes such as gender and race, the field of algorithmic bias has introduced various fairness concepts, along with methodologies to achieve these notions in different contexts. Despite the rapid advancement, not all sectors have embraced these fairness principles to the same extent. One specific sector that merits attention in this regard is insurance. Within the realm of insurance pricing, fairness is defined through a distinct and specialized framework. Consequently, achieving fairness according to established notions does not automatically ensure fair pricing in insurance. In particular, regulators are increasingly emphasizing transparency in pricing algorithms and imposing constraints on insurance companies on the collection and utilization of sensitive consumer attributes. These factors present additional challenges in the implementation of fairness in pricing algorithms. To address these complexities and comply with regulatory demands, we propose an efficient method for constructing fair models that are tailored to the insurance domain, using only privatized sensitive attributes. Notably, our approach ensures statistical guarantees, does not require direct access to sensitive attributes, and adapts to varying transparency requirements, addressing regulatory demands while ensuring fairness in insurance pricing.
♻ ☆ Poisson Midpoint Method for Log Concave Sampling: Beyond the Strong Error Lower Bounds
We study the problem of sampling from strongly log-concave distributions over $\mathbb{R}^d$ using the Poisson midpoint discretization (a variant of the randomized midpoint method) for overdamped/underdamped Langevin dynamics. We prove its convergence in the 2-Wasserstein distance ($W_2$), achieving a cubic speedup in dependence on the target accuracy ($\epsilon$) over the Euler-Maruyama discretization, surpassing existing bounds for randomized midpoint methods. Notably, in the case of underdamped Langevin dynamics, we demonstrate the complexity of $W_2$ convergence is much smaller than the complexity lower bounds for convergence in $L^2$ strong error established in the literature.
♻ ☆ Roll the dice & look before you leap: Going beyond the creative limits of next-token prediction ICML 2025
We design a suite of minimal algorithmic tasks that are a loose abstraction of open-ended real-world tasks. This allows us to cleanly and controllably quantify the creative limits of the present-day language model. Much like real-world tasks that require a creative, far-sighted leap of thought, our tasks require an implicit, open-ended stochastic planning step that either (a) discovers new connections in an abstract knowledge graph (like in wordplay, drawing analogies, or research) or (b) constructs new patterns (like in designing math problems or new proteins). In these tasks, we empirically and conceptually argue how next-token learning is myopic; multi-token approaches, namely teacherless training and diffusion models, comparatively excel in producing diverse and original output. Secondly, to elicit randomness without hurting coherence, we find that injecting noise at the input layer (dubbed seed-conditioning) works surprisingly as well as (and in some conditions, better than) temperature sampling from the output layer. Thus, our work offers a principled, minimal test-bed for analyzing open-ended creative skills, and offers new arguments for going beyond next-token learning and temperature sampling. We make part of the code available under https://github.com/chenwu98/algorithmic-creativity
comment: ICML 2025 (oral)
♻ ☆ SEAL: Towards Safe Autonomous Driving via Skill-Enabled Adversary Learning for Closed-Loop Scenario Generation
Verification and validation of autonomous driving (AD) systems and components is of increasing importance, as such technology increases in real-world prevalence. Safety-critical scenario generation is a key approach to robustify AD policies through closed-loop training. However, existing approaches for scenario generation rely on simplistic objectives, resulting in overly-aggressive or non-reactive adversarial behaviors. To generate diverse adversarial yet realistic scenarios, we propose SEAL, a scenario perturbation approach which leverages learned objective functions and adversarial, human-like skills. SEAL-perturbed scenarios are more realistic than SOTA baselines, leading to improved ego task success across real-world, in-distribution, and out-of-distribution scenarios, of more than 20%. To facilitate future research, we release our code and tools: https://github.com/cmubig/SEAL
comment: Accepted to the IEEE Robotics and Automation Letters (RA-L) on June 28, 2025
♻ ☆ Bypassing LLM Guardrails: An Empirical Analysis of Evasion Attacks against Prompt Injection and Jailbreak Detection Systems
Large Language Models (LLMs) guardrail systems are designed to protect against prompt injection and jailbreak attacks. However, they remain vulnerable to evasion techniques. We demonstrate two approaches for bypassing LLM prompt injection and jailbreak detection systems via traditional character injection methods and algorithmic Adversarial Machine Learning (AML) evasion techniques. Through testing against six prominent protection systems, including Microsoft's Azure Prompt Shield and Meta's Prompt Guard, we show that both methods can be used to evade detection while maintaining adversarial utility achieving in some instances up to 100% evasion success. Furthermore, we demonstrate that adversaries can enhance Attack Success Rates (ASR) against black-box targets by leveraging word importance ranking computed by offline white-box models. Our findings reveal vulnerabilities within current LLM protection mechanisms and highlight the need for more robust guardrail systems.
comment: 14 pages, 5 figures, 11 tables. To be published in LLMSec 2025
♻ ☆ Deep Learning Accelerated Quantum Transport Simulations in Nanoelectronics: From Break Junctions to Field-Effect Transistors
Quantum transport simulations are essential for understanding and designing nanoelectronic devices, yet the long-standing trade-off between accuracy and computational efficiency has limited their practical applications. We present DeePTB-NEGF, an integrated framework combining deep learning tight-binding Hamiltonian prediction with non-equilibrium Green's Function methodology to enable accurate quantum transport simulations in open boundary conditions with 2-3 orders of magnitude acceleration. We demonstrate DeePTB-NEGF through two challenging applications: comprehensive break junction simulations with over $10^4$ snapshots, showing excellent agreement with experimental conductance histograms; and carbon nanotube field-effect transistors (CNT-FET) at experimental dimensions, reproducing measured transfer characteristics for a 41 nm channel CNT-FET ($\sim 8000$ atoms, $3\times10^4$ orbitals) and predicting zero-bias transmission spectra for a 180 nm CNT ($\sim 3\times 10^4$ atoms, $10^5$ orbitals), showcasing the framework's capability for large-scale device simulations. Our systematic studies across varying geometries confirm the necessity of simulating realistic experimental structures for precise predictions. DeePTB-NEGF bridges the longstanding gap between first-principles accuracy and computational efficiency, providing a scalable tool for high-throughput and large-scale quantum transport simulations that enables previously inaccessible nanoscale device investigations.
comment: 10 pages, 4 figures
♻ ☆ EVOLvE: Evaluating and Optimizing LLMs For In-Context Exploration ICML 2025
Despite their success in many domains, large language models (LLMs) remain under-studied in scenarios requiring optimal decision-making under uncertainty. This is crucial as many real-world applications, ranging from personalized recommendations to healthcare interventions, demand that LLMs not only predict but also actively learn to make optimal decisions through exploration. In this work, we measure LLMs' (in)ability to make optimal decisions in bandits, a state-less reinforcement learning setting relevant to many applications. We develop a comprehensive suite of environments, including both context-free and contextual bandits with varying task difficulties, to benchmark LLMs' performance. Motivated by the existence of optimal exploration algorithms, we propose efficient ways to integrate this algorithmic knowledge into LLMs: by providing explicit algorithm-guided support during inference; and through algorithm distillation via in-context demonstrations and fine-tuning, using synthetic data generated from these algorithms. Impressively, these techniques allow us to achieve superior exploration performance with smaller models, surpassing larger models on various tasks. We conducted an extensive ablation study to shed light on various factors, such as task difficulty and data representation, that influence the efficiency of LLM exploration. Additionally, we conduct a rigorous analysis of the LLM's exploration efficiency using the concept of regret, linking its ability to explore to the model size and underlying algorithm.
comment: 28 pages. Published at ICML 2025
♻ ☆ Average Calibration Error: A Differentiable Loss for Improved Reliability in Image Segmentation
Deep neural networks for medical image segmentation often produce overconfident results misaligned with empirical observations. Such miscalibration, challenges their clinical translation. We propose to use marginal L1 average calibration error (mL1-ACE) as a novel auxiliary loss function to improve pixel-wise calibration without compromising segmentation quality. We show that this loss, despite using hard binning, is directly differentiable, bypassing the need for approximate but differentiable surrogate or soft binning approaches. Our work also introduces the concept of dataset reliability histograms which generalises standard reliability diagrams for refined visual assessment of calibration in semantic segmentation aggregated at the dataset level. Using mL1-ACE, we reduce average and maximum calibration error by 45% and 55% respectively, maintaining a Dice score of 87% on the BraTS 2021 dataset. We share our code here: https://github.com/cai4cai/ACE-DLIRIS
comment: Camera ready version as in 10.1007/978-3-031-72114-4_14
♻ ☆ Faster Reinforcement Learning by Freezing Slow States
We study infinite horizon Markov decision processes (MDPs) with "fast-slow" structure, where some state variables evolve rapidly ("fast states") while others change more gradually ("slow states"). This structure commonly arises in practice when decisions must be made at high frequencies over long horizons, and where slowly changing information still plays a critical role in determining optimal actions. Examples include inventory control under slowly changing demand indicators or dynamic pricing with gradually shifting consumer behavior. Modeling the problem at the natural decision frequency leads to MDPs with discount factors close to one, making them computationally challenging. We propose a novel approximation strategy that "freezes" slow states during phases of lower-level planning and subsequently applies value iteration to an auxiliary upper-level MDP that evolves on a slower timescale. Freezing states for short periods of time leads to easier-to-solve lower-level problems, while a slower upper-level timescale allows for a more favorable discount factor. On the theoretical side, we analyze the regret incurred by our frozen-state approach, which leads to simple insights on how to trade off regret versus computational cost. Empirically, we benchmark our new frozen-state methods on three domains, (i) inventory control with fixed order costs, (ii) a gridworld problem with spatial tasks, and (iii) dynamic pricing with reference-price effects. We demonstrate that the new methods produce high-quality policies with significantly less computation, and we show that simply omitting slow states is often a poor heuristic.
comment: 70 pages, 10 figures
♻ ☆ Constructing Extreme Heatwave Storylines with Differentiable Climate Models
Understanding the plausible upper bounds of extreme weather events is essential for risk assessment in a warming climate. Existing methods, based on large ensembles of physics-based models, are often computationally expensive or lack the fidelity needed to simulate rare, high-impact extremes. Here, we present a novel framework that leverages a differentiable hybrid climate model, NeuralGCM, to optimize initial conditions and generate physically consistent worst-case heatwave trajectories. Applied to the 2021 Pacific Northwest heatwave, our method produces heatwave intensity up to 3.7 $^\circ$C above the most extreme member of a 75-member ensemble. These trajectories feature intensified atmospheric blocking and amplified Rossby wave patterns-hallmarks of severe heat events. Our results demonstrate that differentiable climate models can efficiently explore the upper tails of event likelihoods, providing a powerful new approach for constructing targeted storylines of extreme weather under climate change.
♻ ☆ Zero-Shot Cyclic Peptide Design via Composable Geometric Constraints
Cyclic peptides, characterized by geometric constraints absent in linear peptides, offer enhanced biochemical properties, presenting new opportunities to address unmet medical needs. However, designing target-specific cyclic peptides remains underexplored due to limited training data. To bridge the gap, we propose CP-Composer, a novel generative framework that enables zero-shot cyclic peptide generation via composable geometric constraints. Our approach decomposes complex cyclization patterns into unit constraints, which are incorporated into a diffusion model through geometric conditioning on nodes and edges. During training, the model learns from unit constraints and their random combinations in linear peptides, while at inference, novel constraint combinations required for cyclization are imposed as input. Experiments show that our model, despite trained with linear peptides, is capable of generating diverse target-binding cyclic peptides, reaching success rates from 38% to 84% on different cyclization strategies.
♻ ☆ LEXam: Benchmarking Legal Reasoning on 340 Law Exams
Long-form legal reasoning remains a key challenge for large language models (LLMs) in spite of recent advances in test-time scaling. We introduce LEXam, a novel benchmark derived from 340 law exams spanning 116 law school courses across a range of subjects and degree levels. The dataset comprises 4,886 law exam questions in English and German, including 2,841 long-form, open-ended questions and 2,045 multiple-choice questions. Besides reference answers, the open questions are also accompanied by explicit guidance outlining the expected legal reasoning approach such as issue spotting, rule recall, or rule application. Our evaluation on both open-ended and multiple-choice questions present significant challenges for current LLMs; in particular, they notably struggle with open questions that require structured, multi-step legal reasoning. Moreover, our results underscore the effectiveness of the dataset in differentiating between models with varying capabilities. Adopting an LLM-as-a-Judge paradigm with rigorous human expert validation, we demonstrate how model-generated reasoning steps can be evaluated consistently and accurately. Our evaluation setup provides a scalable method to assess legal reasoning quality beyond simple accuracy metrics. Project page: https://lexam-benchmark.github.io/
♻ ☆ DESIGN: Encrypted GNN Inference via Server-Side Input Graph Pruning NeurIPS 2025
Graph Neural Networks (GNNs) have achieved state-of-the-art performance in various graph-based learning tasks. However, enabling privacy-preserving GNNs in encrypted domains, such as under Fully Homomorphic Encryption (FHE), typically incurs substantial computational overhead, rendering real-time and privacy-preserving inference impractical. In this work, we propose DESIGN (EncrypteD GNN Inference via sErver-Side Input Graph pruNing), a novel framework for efficient encrypted GNN inference. DESIGN tackles the critical efficiency limitations of existing FHE GNN approaches, which often overlook input data redundancy and apply uniform computational strategies. Our framework achieves significant performance gains through a hierarchical optimization strategy executed entirely on the server: first, FHE-compatible node importance scores (based on encrypted degree statistics) are computed from the encrypted graph. These scores then guide a homomorphic partitioning process, generating multi-level importance masks directly under FHE. This dynamically generated mask facilitates both input graph pruning (by logically removing unimportant elements) and a novel adaptive polynomial activation scheme, where activation complexity is tailored to node importance levels. Empirical evaluations demonstrate that DESIGN substantially accelerates FHE GNN inference compared to state-of-the-art methods while maintaining competitive model accuracy, presenting a robust solution for secure graph analytics. Our implementation is publicly available at https://github.com/LabRAI/DESIGN.
comment: Under Review in Conference on Neural Information Processing Systems (NeurIPS 2025)
♻ ☆ TKAN: Temporal Kolmogorov-Arnold Networks
Recurrent Neural Networks (RNNs) have revolutionized many areas of machine learning, particularly in natural language and data sequence processing. Long Short-Term Memory (LSTM) has demonstrated its ability to capture long-term dependencies in sequential data. Inspired by the Kolmogorov-Arnold Networks (KANs) a promising alternatives to Multi-Layer Perceptrons (MLPs), we proposed a new neural networks architecture inspired by KAN and the LSTM, the Temporal Kolomogorov-Arnold Networks (TKANs). TKANs combined the strenght of both networks, it is composed of Recurring Kolmogorov-Arnold Networks (RKANs) Layers embedding memory management. This innovation enables us to perform multi-step time series forecasting with enhanced accuracy and efficiency. By addressing the limitations of traditional models in handling complex sequential patterns, the TKAN architecture offers significant potential for advancements in fields requiring more than one step ahead forecasting.
♻ ☆ Low Resource Reconstruction Attacks Through Benign Prompts
The recent advances in generative models such as diffusion models have raised several risks and concerns related to privacy, copyright infringements and data stewardship. To better understand and control the risks, various researchers have created techniques, experiments and attacks that reconstruct images, or part of images, from the training set. While these techniques already establish that data from the training set can be reconstructed, they often rely on high-resources, excess to the training set as well as well-engineered and designed prompts. In this work, we devise a new attack that requires low resources, assumes little to no access to the actual training set, and identifies, seemingly, benign prompts that lead to potentially-risky image reconstruction. This highlights the risk that images might even be reconstructed by an uninformed user and unintentionally. For example, we identified that, with regard to one existing model, the prompt ``blue Unisex T-Shirt'' can generate the face of a real-life human model. Our method builds on an intuition from previous works which leverages domain knowledge and identifies a fundamental vulnerability that stems from the use of scraped data from e-commerce platforms, where templated layouts and images are tied to pattern-like prompts.
♻ ☆ Application of RESNET50 Convolution Neural Network for the Extraction of Optical Parameters in Scattering Media
Estimation of the optical properties of scattering media such as tissue is important in diagnostics as well as in the development of techniques to image deeper. As light penetrates the sample scattering events occur that alter the propagation direction of the photons in a random manner leading degradation of image quality. The distribution of the scattered light does, however, give a measure of the optical properties such as the reduced scattering coefficient and the absorption coefficient. Unfortunately, inverting scattering patterns to recover the optical properties is not simple especially in the regime where the light is partially randomized. Machine learning has been proposed by several authors as a means of recovering these properties from either the back scattered or the transmitted light. In the present paper we train a general purpose convolutional neural network RESNET 50 with simulated data based on Monte Carlo simulations. We show that compared with previous work our approach gives comparable or better reconstruction accuracy with training on a much smaller dataset. Moreover, by training on multiple parameters such as the intensity distribution at multiple planes or the exit angle and spatial distribution one achieves improved performance compared to training on a single input such as the intensity distribution captured at the sample surface. While our approach gives good parameter reconstruction, we identify factors that limit accuracy of the recovered properties, particularly the absorption coefficient. In the light of these limitations, we suggest how the present approach may be enhanced for even better performance.
♻ ☆ Asymptotic regularity of a generalised stochastic Halpern scheme
We provide abstract, general and highly uniform rates of asymptotic regularity for a generalized stochastic Halpern-style iteration, which incorporates a second mapping in the style of a Krasnoselskii-Mann iteration. This iteration is general in two ways: First, it incorporates stochasticity in a completely abstract way rather than fixing a sampling method; secondly, it includes as special cases stochastic versions of various schemes from the optimization literature, including Halpern's iteration as well as a Krasnoselskii-Mann iteration with Tikhonov regularization terms in the sense of Bo\c{t}, Csetnek and Meier. For these specific cases, we in particular obtain linear rates of asymptotic regularity, matching (or improving) the currently best known rates for these iterations in stochastic optimization, and quadratic rates of asymptotic regularity are obtained in the context of inner product spaces for the general iteration. At the end, we briefly sketch how the schemes presented here can be instantiated in the context of reinforcement learning to yield novel methods for Q-learning.
comment: 25 pages
♻ ☆ On the asymptotic behaviour of stochastic processes, with applications to supermartingale convergence, Dvoretzky's approximation theorem, and stochastic quasi-Fejér monotonicity
We prove a novel and general result on the asymptotic behavior of stochastic processes which conform to a certain relaxed supermartingale condition. Our result provides quantitative information in the form of an explicit and effective construction of a rate of convergence for this process, both in mean and almost surely, that is moreover highly uniform in that it only depends on very few data of the surrounding objects involved in the iteration. We then apply this result to derive new quantitative versions of well-known concepts and theorems from stochastic approximation, in particular providing effective rates for a variant of the Robbins-Siegmund theorem, Dvoretzky's convergence theorem, as well as the convergence of stochastic quasi-Fej\'er monotone sequences, the latter of which formulated in a novel and highly general metric context. We utilize the classic and widely studied Robbins-Monro procedure as a template to evaluate our quantitative results and their applicability in greater detail. We conclude by illustrating the breadth of potential further applications with a brief discussion on a variety of other well-known iterative procedures from stochastic approximation. Throughout, we isolate and discuss special cases of our results which allow for the construction of fast, and in particular linear, rates.
comment: 40 pages
♻ ☆ Transformers Can Solve Non-Linear and Non-Markovian Filtering Problems in Continuous Time For Conditionally Gaussian Signals
The use of attention-based deep learning models in stochastic filtering, e.g. transformers and deep Kalman filters, has recently come into focus; however, the potential for these models to solve stochastic filtering problems remains largely unknown. The paper provides an affirmative answer to this open problem in the theoretical foundations of machine learning by showing that a class of continuous-time transformer models, called \textit{filterformers}, can approximately implement the conditional law of a broad class of non-Markovian and conditionally Gaussian signal processes given noisy continuous-time (possibly non-Gaussian) measurements. Our approximation guarantees hold uniformly over sufficiently regular compact subsets of continuous-time paths, where the worst-case 2-Wasserstein distance between the true optimal filter and our deep learning model quantifies the approximation error. Our construction relies on two new customizations of the standard attention mechanism: The first can losslessly adapt to the characteristics of a broad range of paths since we show that the attention mechanism implements bi-Lipschitz embeddings of sufficiently regular sets of paths into low-dimensional Euclidean spaces; thus, it incurs no ``dimension reduction error''. The latter attention mechanism is tailored to the geometry of Gaussian measures in the $2$-Wasserstein space. Our analysis relies on new stability estimates of robust optimal filters in the conditionally Gaussian setting.
♻ ☆ Trinity-RFT: A General-Purpose and Unified Framework for Reinforcement Fine-Tuning of Large Language Models
Trinity-RFT is a general-purpose, unified and easy-to-use framework designed for reinforcement fine-tuning (RFT) of large language models. It is built with a modular and decoupled design, consisting of (1) an RFT-core that unifies and generalizes synchronous/asynchronous, on-policy/off-policy, and online/offline modes of RFT; (2) seamless integration for agent-environment interaction with high efficiency and robustness; and (3) systematic data pipelines optimized for RFT. Trinity-RFT can be easily adapted for diverse application scenarios, and serves as a unified platform for development and research of advanced reinforcement learning paradigms at both macroscopic and microscopic levels. This technical report outlines the vision, features, design and implementations of Trinity-RFT, accompanied by extensive examples, applications and experiments that demonstrate its functionalities and user-friendliness.
comment: This technical report will be continuously updated as the codebase evolves. GitHub: https://github.com/modelscope/Trinity-RFT
♻ ☆ Token-based Audio Inpainting via Discrete Diffusion
Audio inpainting refers to the task of reconstructing missing segments in corrupted audio recordings. While prior approaches-including waveform and spectrogram-based diffusion models-have shown promising results for short gaps, they often degrade in quality when gaps exceed 100 milliseconds (ms). In this work, we introduce a novel inpainting method based on discrete diffusion modeling, which operates over tokenized audio representations produced by a pre-trained audio tokenizer. Our approach models the generative process directly in the discrete latent space, enabling stable and semantically coherent reconstruction of missing audio. We evaluate the method on the MusicNet dataset using both objective and perceptual metrics across gap durations up to 300 ms. We further evaluated our approach on the MTG dataset, extending the gap duration to 500 ms. Experimental results demonstrate that our method achieves competitive or superior performance compared to existing baselines, particularly for longer gaps, offering a robust solution for restoring degraded musical recordings. Audio examples of our proposed method can be found at https://iftach21.github.io/
♻ ☆ Unearthing Gems from Stones: Policy Optimization with Negative Sample Augmentation for LLM Reasoning
Recent advances in reasoning language models have witnessed a paradigm shift from short to long CoT pattern. Given the substantial computational cost of rollouts in long CoT models, maximizing the utility of fixed training datasets becomes crucial. Our analysis reveals that negative responses contain valuable components such as self-reflection and error-correction steps, yet primary existing methods either completely discard negative samples (RFT) or apply equal penalization across all tokens (RL), failing to leverage these potential learning signals. In light of this, we propose Behavior Constrained Policy Gradient with Negative Sample Augmentation (BCPG-NSA), a fine-grained offline RL framework that encompasses three stages: 1) sample segmentation, 2) consensus-based step correctness assessment combining LLM and PRM judgers, and 3) policy optimization with NSA designed to effectively mine positive steps within negative samples. Experimental results show that BCPG-NSA outperforms baselines on several challenging math/coding reasoning benchmarks using the same training dataset, achieving improved sample efficiency and demonstrating robustness and scalability when extended to multiple iterations.
♻ ☆ Concentration of measure for non-linear random matrices with applications to neural networks and non-commutative polynomials
We prove concentration inequalities for several models of non-linear random matrices. As corollaries we obtain estimates for linear spectral statistics of the conjugate kernel of neural networks and non-commutative polynomials in (possibly dependent) random matrices.
comment: Some typos fixed (and some new probably introduced), small editorial changes
♻ ☆ Hyperspherical Variational Autoencoders Using Efficient Spherical Cauchy Distribution
We propose a novel variational autoencoder (VAE) architecture that employs a spherical Cauchy (spCauchy) latent distribution. Unlike traditional Gaussian latent spaces or the widely used von Mises-Fisher (vMF) distribution, spCauchy provides a more natural hyperspherical representation of latent variables, better capturing directional data while maintaining flexibility. Its heavy-tailed nature prevents over-regularization, ensuring efficient latent space utilization while offering a more expressive representation. Additionally, spCauchy circumvents the numerical instabilities inherent to vMF, which arise from computing normalization constants involving Bessel functions. Instead, it enables a fully differentiable and efficient reparameterization trick via M\"obius transformations, allowing for stable and scalable training. The KL divergence can be computed through a rapidly converging power series, eliminating concerns of underflow or overflow associated with evaluation of ratios of hypergeometric functions. These properties make spCauchy a compelling alternative for VAEs, offering both theoretical advantages and practical efficiency in high-dimensional generative modeling.
♻ ☆ DiaTool-DPO: Multi-Turn Direct Preference Optimization for Tool-Augmented Large Language Models SIGDIAL 2025
Tool-Augmented Larage Language Models (TA-LLMs) have shown promise in real-world applications, but face challenges in handling incomplete queries and out-of-scope requests. While existing approaches rely mainly on Supervised Fine-Tuning with expert trajectories, we propose DiaTool-DPO, a novel method that enhances TA-LLM's dialogue capabilities through Direct Preference Optimization. We model TA-LLM interactions as a Markov Decision Process with 5 distinct dialogue states and categorize user queries into 3 types based on their state transition trajectories. We automatically construct paired trajectory datasets of correct and incorrect dialogue flows and introduce a specialized objective loss for dialogue control. Our comprehensive evaluation demonstrates that DiaTool-DPO approaches GPT-4o's performance (94.8% in information gathering, 91% in tool call rejection) with substantial improvements over baseline (44% and 9.6% respectively) while maintaining core functionality. Our approach opens new possibilities for developing TA-LLMs that can handle diverse real-world scenarios without requiring additional expert demonstrations or human labeling.
comment: Accepted to SIGDIAL 2025
♻ ☆ Riemannian Time Warping: Multiple Sequence Alignment in Curved Spaces
Temporal alignment of multiple signals through time warping is crucial in many fields, such as classification within speech recognition or robot motion learning. Almost all related works are limited to data in Euclidean space. Although an attempt was made in 2011 to adapt this concept to unit quaternions, a general extension to Riemannian manifolds remains absent. Given its importance for numerous applications in robotics and beyond, we introduce Riemannian Time Warping (RTW). This novel approach efficiently aligns multiple signals by considering the geometric structure of the Riemannian manifold in which the data is embedded. Extensive experiments on synthetic and real-world data, including tests with an LBR iiwa robot, demonstrate that RTW consistently outperforms state-of-the-art baselines in both averaging and classification tasks.
♻ ☆ A Comprehensive Survey of Direct Preference Optimization: Datasets, Theories, Variants, and Applications
With the rapid advancement of large language models (LLMs), aligning policy models with human preferences has become increasingly critical. Direct Preference Optimization (DPO) has emerged as a promising approach for alignment, acting as an RL-free alternative to Reinforcement Learning from Human Feedback (RLHF). Despite DPO's various advancements and inherent limitations, an in-depth review of these aspects is currently lacking in the literature. In this work, we present a comprehensive review of the challenges and opportunities in DPO, covering theoretical analyses, variants, relevant preference datasets, and applications. Specifically, we categorize recent studies on DPO based on key research questions to provide a thorough understanding of DPO's current landscape. Additionally, we propose several future research directions to offer insights on model alignment for the research community. An updated collection of relevant papers can be found on https://github.com/Mr-Loevan/DPO-Survey.
comment: 45 pages, 12 Figures. Project page: https://github.com/Mr-Loevan/DPO-Survey
♻ ☆ Explaining the Impact of Training on Vision Models via Activation Clustering
This paper introduces Neuro-Activated Vision Explanations (NAVE), a method for extracting and visualizing the internal representations of vision model encoders. By clustering feature activations, NAVE provides insights into learned semantics without fine-tuning. Using object localization, we show that NAVE's concepts align with image semantics. Through extensive experiments, we analyze the impact of training strategies and architectures on encoder representation capabilities. Additionally, we apply NAVE to study training artifacts in vision transformers and reveal how weak training strategies and spurious correlations degrade model performance. Our findings establish NAVE as a valuable tool for post-hoc model inspection and improving transparency in vision models.
♻ ☆ Structuring Radiology Reports: Challenging LLMs with Lightweight Models
Radiology reports are critical for clinical decision-making but often lack a standardized format, limiting both human interpretability and machine learning (ML) applications. While large language models (LLMs) have shown strong capabilities in reformatting clinical text, their high computational requirements, lack of transparency, and data privacy concerns hinder practical deployment. To address these challenges, we explore lightweight encoder-decoder models (<300M parameters)-specifically T5 and BERT2BERT-for structuring radiology reports from the MIMIC-CXR and CheXpert Plus datasets. We benchmark these models against eight open-source LLMs (1B-70B), adapted using prefix prompting, in-context learning (ICL), and low-rank adaptation (LoRA) finetuning. Our best-performing lightweight model outperforms all LLMs adapted using prompt-based techniques on a human-annotated test set. While some LoRA-finetuned LLMs achieve modest gains over the lightweight model on the Findings section (BLEU 6.4%, ROUGE-L 4.8%, BERTScore 3.6%, F1-RadGraph 1.1%, GREEN 3.6%, and F1-SRR-BERT 4.3%), these improvements come at the cost of substantially greater computational resources. For example, LLaMA-3-70B incurred more than 400 times the inference time, cost, and carbon emissions compared to the lightweight model. These results underscore the potential of lightweight, task-specific models as sustainable and privacy-preserving solutions for structuring clinical text in resource-constrained healthcare settings.
♻ ☆ FRUGAL: Memory-Efficient Optimization by Reducing State Overhead for Scalable Training
With the increase in the number of parameters in large language models, the process of pre-training and fine-tuning increasingly demands larger volumes of GPU memory. A significant portion of this memory is typically consumed by the optimizer state. To overcome this challenge, recent approaches such as low-rank adaptation (LoRA (Hu et al., 2021)), low-rank gradient projection (GaLore (Zhao et al., 2024)), and blockwise optimization (BAdam (Luo et al., 2024)) have been proposed. However, in all these algorithms, the $\textit{effective rank of the weight updates remains low-rank}$, which can lead to a substantial loss of information from the gradient. This loss can be critically important, especially during the pre-training stage. In this paper, we introduce $\texttt{FRUGAL}$ ($\textbf{F}$ull-$\textbf{R}$ank $\textbf{U}$pdates with $\textbf{G}$r$\textbf{A}$dient sp$\textbf{L}$itting), a new memory-efficient optimization framework. $\texttt{FRUGAL}$ leverages gradient splitting to perform low-dimensional updates using advanced algorithms (such as Adam), while updates along the remaining directions are executed via state-free methods like SGD or signSGD (Bernstein et al., 2018). Our framework can be integrated with various low-rank update selection techniques, including GaLore and BAdam. We provide theoretical convergence guarantees for our framework when using SGDM for low-dimensional updates and SGD for state-free updates. Additionally, our method consistently outperforms concurrent approaches across various fixed memory budgets, achieving state-of-the-art results in pre-training and fine-tuning tasks while balancing memory efficiency and performance metrics.
♻ ☆ Class-Aware PillarMix: Can Mixed Sample Data Augmentation Enhance 3D Object Detection with Radar Point Clouds? IROS 2025
Due to the significant effort required for data collection and annotation in 3D perception tasks, mixed sample data augmentation (MSDA) has been widely studied to generate diverse training samples by mixing existing data. Recently, many MSDA techniques have been developed for point clouds, but they mainly target LiDAR data, leaving their application to radar point clouds largely unexplored. In this paper, we examine the feasibility of applying existing MSDA methods to radar point clouds and identify several challenges in adapting these techniques. These obstacles stem from the radar's irregular angular distribution, deviations from a single-sensor polar layout in multi-radar setups, and point sparsity. To address these issues, we propose Class-Aware PillarMix (CAPMix), a novel MSDA approach that applies MixUp at the pillar level in 3D point clouds, guided by class labels. Unlike methods that rely a single mix ratio to the entire sample, CAPMix assigns an independent ratio to each pillar, boosting sample diversity. To account for the density of different classes, we use class-specific distributions: for dense objects (e.g., large vehicles), we skew ratios to favor points from another sample, while for sparse objects (e.g., pedestrians), we sample more points from the original. This class-aware mixing retains critical details and enriches each sample with new information, ultimately generating more diverse training data. Experimental results demonstrate that our method not only significantly boosts performance but also outperforms existing MSDA approaches across two datasets (Bosch Street and K-Radar). We believe that this straightforward yet effective approach will spark further investigation into MSDA techniques for radar data.
comment: 8 pages, 6 figures, 4 tables, accepted to 2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS 2025)
♻ ☆ RsGCN: Rescaling Enhances Generalization of GCNs for Solving Scalable Traveling Salesman Problems
Neural traveling salesman problem (TSP) solvers face two critical challenges: poor generalization for scalable TSPs and high training costs. To address these challenges, we propose a new Rescaling Graph Convolutional Network (RsGCN). Focusing on the scale-dependent features (i.e., features varied with problem scales) related to nodes and edges that influence the sensitivity of GCNs to the problem scales, a Rescaling Mechanism in RsGCN enhances the generalization capability by (1) rescaling adjacent nodes to construct a subgraph with a uniform number of adjacent nodes for each node across various scales of TSPs, which stabilizes the graph message aggregation; (2) rescaling subgraph edges to adjust the lengths of subgraph edges to the same magnitude, which maintains numerical consistency. In addition, an efficient training strategy with a mixed-scale dataset and bidirectional loss is used in RsGCN. To fully exploit the heatmaps generated by RsGCN, we design an efficient post-search algorithm termed Re2Opt, in which a reconstruction process based on adaptive weight is incorporated to help avoid local optima. Based on a combined architecture of RsGCN and Re2Opt, our solver achieves remarkable generalization and low training cost: with only 3 epochs of training on the mixed-scale dataset containing instances with up to 100 nodes, it can be generalized successfully to 10K-node instances without any fine-tuning. Extensive experiments demonstrate our state-of-the-art performance across uniform distribution instances of 9 different scales from 20 to 10K nodes and 78 real-world instances from TSPLIB, while requiring the fewest learnable parameters and training epochs among neural competitors.
♻ ☆ Quality over Quantity: An Effective Large-Scale Data Reduction Strategy Based on Pointwise V-Information
In order to increase the effectiveness of model training, data reduction is essential to data-centric AI. It does this by locating the most instructive examples in massive datasets. To increase data quality and training efficiency, the main difficulty is to choose the best examples rather than the complete datasets. In this paper, we propose an effective data reduction strategy based on Pointwise -Information (PVI). To enable a static method, we first use PVI to quantify instance difficulty and remove instances with low difficulty. Experiments show that the classifier performance is maintained with only a 0.0001% to 0.76% reduction in accuracy when 10%-30% of the data is removed. Second, we train the classifiers using a progressive learning strategy on examples sorted by increasing PVI, accelerating convergence and achieving a 0.8% accuracy gain over conventional training. Our findings imply that training a classifier on the chosen optimal subset may improve model performance and increase training efficiency when combined with an efficient data reduction strategy. Furthermore, we have adapted the PVI framework, which was previously limited to English datasets, to a variety of Chinese NLP tasks and base models, yielding insightful results for faster training and cross-lingual data reduction. The codes are released at https://github.com/zhouwenchi/DatasetReductionStrategy.
♻ ☆ A Vector-Quantized Foundation Model for Patient Behavior Monitoring
Foundation models have achieved remarkable success across various domains, yet their adoption in healthcare remains limited. While significant advances have been made in medical imaging, genetic biomarkers, and time series from electronic health records, the potential of foundation models for patient behavior monitoring through personal digital devices remains underexplored. The data generated by these devices are inherently heterogeneous, multisource, and often exhibit high rates of missing data, posing unique challenges. This paper introduces a novel foundation model based on a modified vector quantized variational autoencoder, specifically designed to process real-world data from smartphones and wearable devices. We leveraged the discrete latent representation of this model to effectively perform two downstream tasks, suicide risk assessment and emotional state prediction, on different held-out clinical cohorts without the need of fine-tuning. We also highlight the existence of a trade-off between discrete and continuous latent structures, suggesting that hybrid models may be optimal for balancing accuracy across various supervised and unsupervised tasks.
comment: 10 pages (32 with references and supplementary material). Submitted to IEEE Journal of Biomedical and Health Informatics
♻ ☆ Self-attentive Transformer for Fast and Accurate Postprocessing of Temperature and Wind Speed Forecasts
Current postprocessing techniques often require separate models for each lead time and disregard possible inter-ensemble relationships by either correcting each member separately or by employing distributional approaches. In this work, we tackle these shortcomings with an innovative, fast and accurate Transformer which postprocesses each ensemble member individually while allowing information exchange across variables, spatial dimensions and lead times by means of multi-headed self-attention. Weather forecasts are postprocessed over 20 lead times simultaneously while including up to fifteen meteorological predictors. We use the EUPPBench dataset for training which contains ensemble predictions from the European Center for Medium-range Weather Forecasts' integrated forecasting system alongside corresponding observations. The work presented here is the first to postprocess the ten and one hundred-meter wind speed forecasts within this benchmark dataset, while also correcting two-meter temperature. Our approach significantly improves the original forecasts, as measured by the CRPS, with 16.5\% for two-meter temperature, 10\% for ten-meter wind speed and 9\% for one hundred-meter wind speed, outperforming a classical member-by-member approach employed as a competitive benchmark. Furthermore, being up to six times faster, it fulfills the demand for rapid operational weather forecasts in various downstream applications, including renewable energy forecasting.
comment: 23 pages, 9 figures, Accepted for publication in Artificial Intelligence for the Earth Systems (AIES)
♻ ☆ IPAD: Inverse Prompt for AI Detection -- A Robust and Explainable LLM-Generated Text Detector
Large Language Models (LLMs) have attained human-level fluency in text generation, which complicates the distinction between human-written and LLM-generated texts. This increases the risk of misuse and highlights the need for reliable detectors. Yet, existing detectors exhibit poor robustness on out-of-distribution (OOD) data and attacked data, which is critical for real-world scenarios. Also, they struggle to provide interpretable evidence to support their decisions, thus undermining the reliability. In light of these challenges, we propose IPAD (Inverse Prompt for AI Detection), a novel framework consisting of a Prompt Inverter that identifies predicted prompts that could have generated the input text, and two Distinguishers that examine the probability that the input texts align with the predicted prompts. Empirical evaluations demonstrate that IPAD outperforms the strongest baselines by 9.05% (Average Recall) on in-distribution data, 12.93% (AUROC) on out-of-distribution (OOD) data, and 5.48% (AUROC) on attacked data. IPAD also performs robustly on structured datasets. Furthermore, an interpretability assessment is conducted to illustrate that IPAD enhances the AI detection trustworthiness by allowing users to directly examine the decision-making evidence, which provides interpretable support for its state-of-the-art detection results.
♻ ☆ On the Learning with Augmented Class via Forests IJCAI 2025
Decision trees and forests have achieved successes in various real applications, most working with all testing classes known in training data. In this work, we focus on learning with augmented class via forests, where an augmented class may appear in testing data yet not in training data. We incorporate information of augmented class into trees' splitting, that is, augmented Gini impurity, a new splitting criterion is introduced to exploit some unlabeled data from testing distribution. We then develop the Learning with Augmented Class via Forests (short for LACForest) approach, which constructs shallow forests according to the augmented Gini impurity and then splits forests with pseudo-labeled augmented instances for better performance. We also develop deep neural forests via an optimization objective based on our augmented Gini impurity, which essentially utilizes the representation power of neural networks for forests. Theoretically, we present the convergence analysis for our augmented Gini impurity, and we finally conduct experiments to evaluate our approaches. The code is available at https://github.com/nju-xuf/LACForest.
comment: Accepted by IJCAI 2025
♻ ☆ Efficient Deployment of Vision-Language Models on Mobile Devices: A Case Study on OnePlus 13R
Vision-Language Models (VLMs) offer promising capabilities for mobile devices, but their deployment faces significant challenges due to computational limitations and energy inefficiency, especially for real-time applications. This study provides a comprehensive survey of deployment frameworks for VLMs on mobile devices, evaluating llama.cpp, MLC-Imp, and mllm in the context of running LLaVA-1.5 7B, MobileVLM-3B, and Imp-v1.5 3B as representative workloads on a OnePlus 13R. Each deployment framework was evaluated on the OnePlus 13R while running VLMs, with measurements covering CPU, GPU, and NPU utilization, temperature, inference time, power consumption, and user experience. Benchmarking revealed critical performance bottlenecks across frameworks: CPU resources were consistently over-utilized during token generation, while GPU and NPU accelerators were largely unused. When the GPU was used, primarily for image feature extraction, it was saturated, leading to degraded device responsiveness. The study contributes framework-level benchmarks, practical profiling tools, and an in-depth analysis of hardware utilization bottlenecks, highlighting the consistent overuse of CPUs and the ineffective or unstable use of GPUs and NPUs in current deployment frameworks.
♻ ☆ First-ish Order Methods: Hessian-aware Scalings of Gradient Descent
Gradient descent is the primary workhorse for optimizing large-scale problems in machine learning. However, its performance is highly sensitive to the choice of the learning rate. A key limitation of gradient descent is its lack of natural scaling, which often necessitates expensive line searches or heuristic tuning to determine an appropriate step size. In this paper, we address this limitation by incorporating Hessian information to scale the gradient direction. By accounting for the curvature of the function along the gradient, our adaptive, Hessian-aware scaling method ensures a local unit step size guarantee, even in nonconvex settings. Near a local minimum that satisfies the second-order sufficient conditions, our approach achieves linear convergence with a unit step size. We show that our method converges globally under a significantly weaker version of the standard Lipschitz gradient smoothness assumption. Even when Hessian information is inexact, the local unit step size guarantee and global convergence properties remain valid under mild conditions. Finally, we validate our theoretical results empirically on a range of convex and nonconvex machine learning tasks, showcasing the effectiveness of the approach.
♻ ☆ Integrated Gradient Correlation: a Dataset-wise Attribution Method
Attribution methods are primarily designed to study input component contributions to individual model predictions. However, some research applications require a summary of attribution patterns across the entire dataset to facilitate the interpretability of the scrutinized models at a task-level rather than an instance-level. It specifically applies when the localization of important input information is supposed to be stable for a specific problem but remains unidentified among numerous components. In this paper, we present a dataset-wise attribution method called Integrated Gradient Correlation (IGC) that enables region-specific analysis by a direct summation over associated components, and further relates the sum of all attributions to a model prediction score (correlation). We demonstrate IGC on synthetic data and fMRI neural signals (NSD dataset) with the study of the representation of image features in the brain and the estimation of the visual receptive field of neural populations. The resulting IGC attributions reveal selective patterns, coherent with respective model objectives.
comment: 16 pages, 6 figures, source code at https://github.com/plelievre/int_grad_corr
♻ ☆ STRAP: Spatial-Temporal Risk-Attentive Vehicle Trajectory Prediction for Autonomous Driving SC 2025
Accurate vehicle trajectory prediction is essential for ensuring safety and efficiency in fully autonomous driving systems. While existing methods primarily focus on modeling observed motion patterns and interactions with other vehicles, they often neglect the potential risks posed by the uncertain or aggressive behaviors of surrounding vehicles. In this paper, we propose a novel spatial-temporal risk-attentive trajectory prediction framework that incorporates a risk potential field to assess perceived risks arising from behaviors of nearby vehicles. The framework leverages a spatial-temporal encoder and a risk-attentive feature fusion decoder to embed the risk potential field into the extracted spatial-temporal feature representations for trajectory prediction. A risk-scaled loss function is further designed to improve the prediction accuracy of high-risk scenarios, such as short relative spacing. Experiments on the widely used NGSIM and HighD datasets demonstrate that our method reduces average prediction errors by 4.8% and 31.2% respectively compared to state-of-the-art approaches, especially in high-risk scenarios. The proposed framework provides interpretable, risk-aware predictions, contributing to more robust decision-making for autonomous driving systems.
comment: 6 pages, 3 figures, accepted at ITSC 2025
♻ ☆ Collaboration Promotes Group Resilience in Multi-Agent AI
To effectively operate in various dynamic scenarios, RL agents must be resilient to unexpected changes in their environment. Previous work on this form of resilience has focused on single-agent settings. In this work, we introduce and formalize a multi-agent variant of resilience, which we term group resilience. We further hypothesize that collaboration with other agents is key to achieving group resilience; collaborating agents adapt better to environmental perturbations in multi-agent reinforcement learning (MARL) settings. We test our hypothesis empirically by evaluating different collaboration protocols and examining their effect on group resilience. Our experiments show that all the examined collaborative approaches achieve higher group resilience than their non-collaborative counterparts.
comment: RLC 2025
♻ ☆ Defense-as-a-Service: Black-box Shielding against Backdoored Graph Models
With the trend of large graph learning models, business owners tend to employ a model provided by a third party to deliver business services to users. However, these models might be backdoored, and malicious users can submit trigger-embedded inputs to manipulate the model predictions. Current graph backdoor defenses have several limitations: 1) depending on model-related details, 2) requiring additional model fine-tuning, and 3) relying upon extra explainability tools, all of which are infeasible under stringent privacy policies. To address those limitations, we propose GraphProt, which allows resource-constrained business owners to rely on third parties to avoid backdoor attacks on GNN-based graph classifiers. Our GraphProt is model-agnostic and only relies on the input graph. The key insight is to leverage subgraph information for prediction, thereby mitigating backdoor effects induced by triggers. GraphProt comprises two components: clustering-based trigger elimination and robust subgraph ensemble. Specifically, we first propose feature-topology clustering that aims to remove most of the anomalous subgraphs (triggers). Moreover, we design subgraph sampling strategies based on feature-topology clustering to build a robust classifier via majority vote. Experimental results across three backdoor attacks and six benchmark datasets demonstrate that GraphProt significantly reduces the backdoor attack success rate while preserving the model accuracy on regular graph classification tasks.
comment: We have to add a rigorous mathematical proof to the thesis proposal, and the process of the current proposal is not rigorous enough
♻ ☆ Teaching LLM to Reason: Reinforcement Learning from Algorithmic Problems without Code
Enhancing reasoning capabilities remains a central focus in the LLM reasearch community. A promising direction involves requiring models to simulate code execution step-by-step to derive outputs for given inputs. However, as code is often designed for large-scale systems, direct application leads to over-reliance on complex data structures and algorithms, even for simple cases, resulting in overfitting to algorithmic patterns rather than core reasoning structures. To address this, we propose TeaR, which aims at teaching LLMs to reason better. TeaR leverages careful data curation and reinforcement learning to guide models in discovering optimal reasoning paths through code-related tasks, thereby improving general reasoning abilities. We conduct extensive experiments using two base models and three long-CoT distillation models, with model sizes ranging from 1.5 billion to 32 billion parameters, and across 17 benchmarks spanning Math, Knowledge, Code, and Logical Reasoning. The results consistently show significant performance improvements. Notably, TeaR achieves a 35.9% improvement on Qwen2.5-7B and 5.9% on R1-Distilled-7B.
♻ ☆ Two-cluster test
Cluster analysis is a fundamental research issue in statistics and machine learning. In many modern clustering methods, we need to determine whether two subsets of samples come from the same cluster. Since these subsets are usually generated by certain clustering procedures, the deployment of classic two-sample tests in this context would yield extremely smaller p-values, leading to inflated Type-I error rate. To overcome this bias, we formally introduce the two-cluster test issue and argue that it is a totally different significance testing issue from conventional two-sample test. Meanwhile, we present a new method based on the boundary points between two subsets to derive an analytical p-value for the purpose of significance quantification. Experiments on both synthetic and real data sets show that the proposed test is able to significantly reduce the Type-I error rate, in comparison with several classic two-sample testing methods. More importantly, the practical usage of such two-cluster test is further verified through its applications in tree-based interpretable clustering and significance-based hierarchical clustering.
♻ ☆ Predicting Graph Structure via Adapted Flux Balance Analysis
Many dynamic processes such as telecommunication and transport networks can be described through discrete time series of graphs. Modelling the dynamics of such time series enables prediction of graph structure at future time steps, which can be used in applications such as detection of anomalies. Existing approaches for graph prediction have limitations such as assuming that the vertices do not to change between consecutive graphs. To address this, we propose to exploit time series prediction methods in combination with an adapted form of flux balance analysis (FBA), a linear programming method originating from biochemistry. FBA is adapted to incorporate various constraints applicable to the scenario of growing graphs. Empirical evaluations on synthetic datasets (constructed via Preferential Attachment model) and real datasets (UCI Message, HePH, Facebook, Bitcoin) demonstrate the efficacy of the proposed approach.
comment: extended and revised version of arXiv:2401.04280
♻ ☆ DeepGesture: A conversational gesture synthesis system based on emotions and semantics
Along with the explosion of large language models, improvements in speech synthesis, advancements in hardware, and the evolution of computer graphics, the current bottleneck in creating digital humans lies in generating character movements that correspond naturally to text or speech inputs. In this work, we present DeepGesture, a diffusion-based gesture synthesis framework for generating expressive co-speech gestures conditioned on multimodal signals - text, speech, emotion, and seed motion. Built upon the DiffuseStyleGesture model, DeepGesture introduces novel architectural enhancements that improve semantic alignment and emotional expressiveness in generated gestures. Specifically, we integrate fast text transcriptions as semantic conditioning and implement emotion-guided classifier-free diffusion to support controllable gesture generation across affective states. To visualize results, we implement a full rendering pipeline in Unity based on BVH output from the model. Evaluation on the ZeroEGGS dataset shows that DeepGesture produces gestures with improved human-likeness and contextual appropriateness. Our system supports interpolation between emotional states and demonstrates generalization to out-of-distribution speech, including synthetic voices - marking a step forward toward fully multimodal, emotionally aware digital humans. Project page: https://deepgesture.github.io
comment: Project page: https://deepgesture.github.io
♻ ☆ EVALOOP: Assessing LLM Robustness in Programming from a Self-consistency Perspective
Assessing the programming capabilities of Large Language Models (LLMs) is crucial for their effective use in software engineering. Current evaluations, however, predominantly measure the accuracy of generated code on static benchmarks, neglecting the critical aspect of model robustness during programming tasks. While adversarial attacks offer insights on model robustness, their effectiveness is limited and evaluation could be constrained. Current adversarial attack methods for robustness evaluation yield inconsistent results, struggling to provide a unified evaluation across different LLMs. We introduce EVALOOP, a novel assessment framework that evaluate the robustness from a self-consistency perspective, i.e., leveraging the natural duality inherent in popular software engineering tasks, e.g., code generation and code summarization. EVALOOP initiates a self-contained feedback loop: an LLM generates output (e.g., code) from an input (e.g., natural language specification), and then use the generated output as the input to produce a new output (e.g., summarizes that code into a new specification). EVALOOP repeats the process to assess the effectiveness of EVALOOP in each loop. This cyclical strategy intrinsically evaluates robustness without rely on any external attack setups, providing a unified metric to evaluate LLMs' robustness in programming. We evaluate 16 prominent LLMs (e.g., GPT-4.1, O4-mini) on EVALOOP and found that EVALOOP typically induces a 5.01%-19.31% absolute drop in pass@1 performance within ten loops. Intriguingly, robustness does not always align with initial performance (i.e., one-time query); for instance, GPT-3.5-Turbo, despite superior initial code generation compared to DeepSeek-V2, demonstrated lower robustness over repeated evaluation loop.
comment: 20 pages, 11 figures
♻ ☆ Towards Efficient Quantity Retrieval from Text:An Approach via Description Parsing and Weak Supervision
Quantitative facts are continually generated by companies and governments, supporting data-driven decision-making. While common facts are structured, many long-tail quantitative facts remain buried in unstructured documents, making them difficult to access. We propose the task of Quantity Retrieval: given a description of a quantitative fact, the system returns the relevant value and supporting evidence. Understanding quantity semantics in context is essential for this task. We introduce a framework based on description parsing that converts text into structured (description, quantity) pairs for effective retrieval. To improve learning, we construct a large paraphrase dataset using weak supervision based on quantity co-occurrence. We evaluate our approach on a large corpus of financial annual reports and a newly annotated quantity description dataset. Our method significantly improves top-1 retrieval accuracy from 30.98 percent to 64.66 percent.
comment: Extended version of the paper accepted in DEXA 2025
♻ ☆ Exploring Sparse Adapters for Scalable Merging of Parameter Efficient Experts
Merging parameter-efficient task experts has recently gained growing attention as a way to build modular architectures that can be rapidly adapted on the fly for specific downstream tasks, without requiring additional fine-tuning. Typically, LoRA serves as the foundational building block of such parameter-efficient modular architectures, leveraging low-rank weight structures to reduce the number of trainable parameters. In this paper, we study the properties of sparse adapters, which train only a subset of weights in the base neural network, as potential building blocks of modular architectures. First, we propose a simple method for training highly effective sparse adapters, which is conceptually simpler than existing methods in the literature and surprisingly outperforms both LoRA and full fine-tuning in our setting. Next, we investigate the merging properties of these sparse adapters by merging adapters for up to 20 natural language processing tasks, thus scaling beyond what is usually studied in the literature. Our findings demonstrate that sparse adapters yield superior in-distribution performance post-merging compared to LoRA or full model merging. Achieving strong held-out performance remains a challenge for all methods considered.
Multimedia 11
☆ Chat with AI: The Surprising Turn of Real-time Video Communication from Human to AI
AI Video Chat emerges as a new paradigm for Real-time Communication (RTC), where one peer is not a human, but a Multimodal Large Language Model (MLLM). This makes interaction between humans and AI more intuitive, as if chatting face-to-face with a real person. However, this poses significant challenges to latency, because the MLLM inference takes up most of the response time, leaving very little time for video streaming. Due to network uncertainty and instability, transmission latency becomes a critical bottleneck preventing AI from being like a real person. To address this, we propose Artic, an AI-oriented Real-time Communication framework, exploring the network requirement shift from "humans watching video" to "AI understanding video". To reduce bitrate dramatically while maintaining MLLM accuracy, we propose Context-Aware Video Streaming that recognizes the importance of each video region for chat and allocates bitrate almost exclusively to chat-important regions. To avoid packet retransmission, we propose Loss-Resilient Adaptive Frame Rate that leverages previous frames to substitute for lost/delayed frames while avoiding bitrate waste. To evaluate the impact of video streaming quality on MLLM accuracy, we build the first benchmark, named Degraded Video Understanding Benchmark (DeViBench). Finally, we discuss some open questions and ongoing solutions for AI Video Chat.
☆ An Empirical Evaluation of AI-Powered Non-Player Characters' Perceived Realism and Performance in Virtual Reality Environments
Advancements in artificial intelligence (AI) have significantly enhanced the realism and interactivity of non-player characters (NPCs) in virtual reality (VR), creating more engaging and believable user experiences. This paper evaluates AI-driven NPCs within a VR interrogation simulator, focusing on their perceived realism, usability, and system performance. The simulator features two AI-powered NPCs, a suspect, and a partner, using GPT-4 Turbo to engage participants in a scenario to determine the suspect's guilt or innocence. A user study with 18 participants assessed the system using the System Usability Scale (SUS), Game Experience Questionnaire (GEQ), and a Virtual Agent Believability Questionnaire, alongside latency measurements for speech-to-text (STT), text-to-speech (TTS), OpenAI GPT-4 Turbo, and overall (cycle) latency. Results showed an average cycle latency of 7 seconds, influenced by the increasing conversational context. Believability scored 6.67 out of 10, with high ratings in behavior, social relationships, and intelligence but moderate scores in emotion and personality. The system achieved a SUS score of 79.44, indicating good usability. These findings demonstrate the potential of large language models to improve NPC realism and interaction in VR while highlighting challenges in reducing system latency and enhancing emotional depth. This research contributes to the development of more sophisticated AI-driven NPCs, revealing the need for performance optimization to achieve increasingly immersive virtual experiences.
☆ RAPNet: A Receptive-Field Adaptive Convolutional Neural Network for Pansharpening
Pansharpening refers to the process of integrating a high resolution panchromatic (PAN) image with a lower resolution multispectral (MS) image to generate a fused product, which is pivotal in remote sensing. Despite the effectiveness of CNNs in addressing this challenge, they are inherently constrained by the uniform application of convolutional kernels across all spatial positions, overlooking local content variations. To overcome this issue, we introduce RAPNet, a new architecture that leverages content-adaptive convolution. At its core, RAPNet employs the Receptive-field Adaptive Pansharpening Convolution (RAPConv), designed to produce spatially adaptive kernels responsive to local feature context, thereby enhancing the precision of spatial detail extraction. Additionally, the network integrates the Pansharpening Dynamic Feature Fusion (PAN-DFF) module, which incorporates an attention mechanism to achieve an optimal balance between spatial detail enhancement and spectral fidelity. Comprehensive evaluations on publicly available datasets confirm that RAPNet delivers superior performance compared to existing approaches, as demonstrated by both quantitative metrics and qualitative assessments. Ablation analyses further substantiate the effectiveness of the proposed adaptive components.
comment: To appear in the proceedings of the 6th International Conference on Artificial Intelligence and Electromechanical Automation (AIEA 2025). 5 pages, 6 figures
☆ Text-to-Remote-Sensing-Image Retrieval beyond RGB Sources
Retrieving relevant imagery from vast satellite archives is crucial for applications like disaster response and long-term climate monitoring. However, most text-to-image retrieval systems are limited to RGB data, failing to exploit the unique physical information captured by other sensors, such as the all-weather structural sensitivity of Synthetic Aperture Radar (SAR) or the spectral signatures in optical multispectral data. To bridge this gap, we introduce CrisisLandMark, a new large-scale corpus of over 647,000 Sentinel-1 SAR and Sentinel-2 multispectral images paired with structured textual annotations for land cover, land use, and crisis events harmonized from authoritative land cover systems (CORINE and Dynamic World) and crisis-specific sources. We then present CLOSP (Contrastive Language Optical SAR Pretraining), a novel framework that uses text as a bridge to align unpaired optical and SAR images into a unified embedding space. Our experiments show that CLOSP achieves a new state-of-the-art, improving retrieval nDGC by 54% over existing models. Additionally, we find that the unified training strategy overcomes the inherent difficulty of interpreting SAR imagery by transferring rich semantic knowledge from the optical domain with indirect interaction. Furthermore, GeoCLOSP, which integrates geographic coordinates into our framework, creates a powerful trade-off between generality and specificity: while the CLOSP excels at general semantic tasks, the GeoCLOSP becomes a specialized expert for retrieving location-dependent crisis events and rare geographic features. This work highlights that the integration of diverse sensor data and geographic context is essential for unlocking the full potential of remote sensing archives.
☆ DualDub: Video-to-Soundtrack Generation via Joint Speech and Background Audio Synthesis
While recent video-to-audio (V2A) models can generate realistic background audio from visual input, they largely overlook speech, an essential part of many video soundtracks. This paper proposes a new task, video-to-soundtrack (V2ST) generation, which aims to jointly produce synchronized background audio and speech within a unified framework. To tackle V2ST, we introduce DualDub, a unified framework built on a multimodal language model that integrates a multimodal encoder, a cross-modal aligner, and dual decoding heads for simultaneous background audio and speech generation. Specifically, our proposed cross-modal aligner employs causal and non-causal attention mechanisms to improve synchronization and acoustic harmony. Besides, to handle data scarcity, we design a curriculum learning strategy that progressively builds the multimodal capability. Finally, we introduce DualBench, the first benchmark for V2ST evaluation with a carefully curated test set and comprehensive metrics. Experimental results demonstrate that DualDub achieves state-of-the-art performance, generating high-quality and well-synchronized soundtracks with both speech and background audio.
☆ LayLens: Improving Deepfake Understanding through Simplified Explanations
This demonstration paper presents $\mathbf{LayLens}$, a tool aimed to make deepfake understanding easier for users of all educational backgrounds. While prior works often rely on outputs containing technical jargon, LayLens bridges the gap between model reasoning and human understanding through a three-stage pipeline: (1) explainable deepfake detection using a state-of-the-art forgery localization model, (2) natural language simplification of technical explanations using a vision-language model, and (3) visual reconstruction of a plausible original image via guided image editing. The interface presents both technical and layperson-friendly explanations in addition to a side-by-side comparison of the uploaded and reconstructed images. A user study with 15 participants shows that simplified explanations significantly improve clarity and reduce cognitive load, with most users expressing increased confidence in identifying deepfakes. LayLens offers a step toward transparent, trustworthy, and user-centric deepfake forensics.
☆ ESG-Net: Event-Aware Semantic Guided Network for Dense Audio-Visual Event Localization
Dense audio-visual event localization (DAVE) aims to identify event categories and locate the temporal boundaries in untrimmed videos. Most studies only employ event-related semantic constraints on the final outputs, lacking cross-modal semantic bridging in intermediate layers. This causes modality semantic gap for further fusion, making it difficult to distinguish between event-related content and irrelevant background content. Moreover, they rarely consider the correlations between events, which limits the model to infer concurrent events among complex scenarios. In this paper, we incorporate multi-stage semantic guidance and multi-event relationship modeling, which respectively enable hierarchical semantic understanding of audio-visual events and adaptive extraction of event dependencies, thereby better focusing on event-related information. Specifically, our eventaware semantic guided network (ESG-Net) includes a early semantics interaction (ESI) module and a mixture of dependency experts (MoDE) module. ESI applys multi-stage semantic guidance to explicitly constrain the model in learning semantic information through multi-modal early fusion and several classification loss functions, ensuring hierarchical understanding of event-related content. MoDE promotes the extraction of multi-event dependencies through multiple serial mixture of experts with adaptive weight allocation. Extensive experiments demonstrate that our method significantly surpasses the state-of-the-art methods, while greatly reducing parameters and computational load. Our code will be released on https://github.com/uchiha99999/ESG-Net.
☆ MultiVox: Benchmarking Voice Assistants for Multimodal Interactions
The rapid progress of Large Language Models (LLMs) has empowered omni models to act as voice assistants capable of understanding spoken dialogues. These models can process multimodal inputs beyond text, such as speech and visual data, enabling more context-aware interactions. However, current benchmarks fall short in comprehensively evaluating how well these models generate context-aware responses, particularly when it comes to implicitly understanding fine-grained speech characteristics, such as pitch, emotion, timbre, and volume or the environmental acoustic context such as background sounds. Additionally, they inadequately assess the ability of models to align paralinguistic cues with complementary visual signals to inform their responses. To address these gaps, we introduce MultiVox, the first omni voice assistant benchmark designed to evaluate the ability of voice assistants to integrate spoken and visual cues including paralinguistic speech features for truly multimodal understanding. Specifically, MultiVox includes 1000 human-annotated and recorded speech dialogues that encompass diverse paralinguistic features and a range of visual cues such as images and videos. Our evaluation on 9 state-of-the-art models reveals that, although humans excel at these tasks, current models consistently struggle to produce contextually grounded responses.
comment: Work In Progress
♻ ☆ Unsupervised Ego- and Exo-centric Dense Procedural Activity Captioning via Gaze Consensus Adaptation ACM MM 2025
Even from an early age, humans naturally adapt between exocentric (Exo) and egocentric (Ego) perspectives to understand daily procedural activities. Inspired by this cognitive ability, we propose a novel Unsupervised Ego-Exo Dense Procedural Activity Captioning (UE$^{2}$DPAC) task, which aims to transfer knowledge from the labeled source view to predict the time segments and descriptions of action sequences for the target view without annotations. Despite previous works endeavoring to address the fully-supervised single-view or cross-view dense video captioning, they lapse in the proposed task due to the significant inter-view gap caused by temporal misalignment and irrelevant object interference. Hence, we propose a Gaze Consensus-guided Ego-Exo Adaptation Network (GCEAN) that injects the gaze information into the learned representations for the fine-grained Ego-Exo alignment. Specifically, we propose a Score-based Adversarial Learning Module (SALM) that incorporates a discriminative scoring network and compares the scores of distinct views to learn unified view-invariant representations from a global level. Then, the Gaze Consensus Construction Module (GCCM) utilizes the gaze to progressively calibrate the learned representations to highlight the regions of interest and extract the corresponding temporal contexts. Moreover, we adopt hierarchical gaze-guided consistency losses to construct gaze consensus for the explicit temporal and spatial adaptation between the source and target views. To support our research, we propose a new EgoMe-UE$^{2}$DPAC benchmark, and extensive experiments demonstrate the effectiveness of our method, which outperforms many related methods by a large margin. Code is available at https://github.com/ZhaofengSHI/GCEAN.
comment: ACM International Conference on Multimedia(ACM MM 2025)
♻ ☆ AGAV-Rater: Adapting Large Multimodal Model for AI-Generated Audio-Visual Quality Assessment
Many video-to-audio (VTA) methods have been proposed for dubbing silent AI-generated videos. An efficient quality assessment method for AI-generated audio-visual content (AGAV) is crucial for ensuring audio-visual quality. Existing audio-visual quality assessment methods struggle with unique distortions in AGAVs, such as unrealistic and inconsistent elements. To address this, we introduce AGAVQA-3k, the first large-scale AGAV quality assessment dataset, comprising $3,382$ AGAVs from $16$ VTA methods. AGAVQA-3k includes two subsets: AGAVQA-MOS, which provides multi-dimensional scores for audio quality, content consistency, and overall quality, and AGAVQA-Pair, designed for optimal AGAV pair selection. We further propose AGAV-Rater, a LMM-based model that can score AGAVs, as well as audio and music generated from text, across multiple dimensions, and selects the best AGAV generated by VTA methods to present to the user. AGAV-Rater achieves state-of-the-art performance on AGAVQA-3k, Text-to-Audio, and Text-to-Music datasets. Subjective tests also confirm that AGAV-Rater enhances VTA performance and user experience. The dataset and code is available at https://github.com/charlotte9524/AGAV-Rater.
♻ ☆ A Survey on Speech Deepfake Detection
The availability of smart devices leads to an exponential increase in multimedia content. However, advancements in deep learning have also enabled the creation of highly sophisticated Deepfake content, including speech Deepfakes, which pose a serious threat by generating realistic voices and spreading misinformation. To combat this, numerous challenges have been organized to advance speech Deepfake detection techniques. In this survey, we systematically analyze more than 200 papers published up to March 2024. We provide a comprehensive review of each component in the detection pipeline, including model architectures, optimization techniques, generalizability, evaluation metrics, performance comparisons, available datasets, and open source availability. For each aspect, we assess recent progress and discuss ongoing challenges. In addition, we explore emerging topics such as partial Deepfake detection, cross-dataset evaluation, and defences against adversarial attacks, while suggesting promising research directions. This survey not only identifies the current state of the art to establish strong baselines for future experiments but also offers clear guidance for researchers aiming to enhance speech Deepfake detection systems.
comment: 38 pages. This paper has been accepted by ACM Computing Surveys
Computation and Language 47
☆ TinyTroupe: An LLM-powered Multiagent Persona Simulation Toolkit
Recent advances in Large Language Models (LLM) have led to a new class of autonomous agents, renewing and expanding interest in the area. LLM-powered Multiagent Systems (MAS) have thus emerged, both for assistive and simulation purposes, yet tools for realistic human behavior simulation -- with its distinctive challenges and opportunities -- remain underdeveloped. Existing MAS libraries and tools lack fine-grained persona specifications, population sampling facilities, experimentation support, and integrated validation, among other key capabilities, limiting their utility for behavioral studies, social simulation, and related applications. To address these deficiencies, in this work we introduce TinyTroupe, a simulation toolkit enabling detailed persona definitions (e.g., nationality, age, occupation, personality, beliefs, behaviors) and programmatic control via numerous LLM-driven mechanisms. This allows for the concise formulation of behavioral problems of practical interest, either at the individual or group level, and provides effective means for their solution. TinyTroupe's components are presented using representative working examples, such as brainstorming and market research sessions, thereby simultaneously clarifying their purpose and demonstrating their usefulness. Quantitative and qualitative evaluations of selected aspects are also provided, highlighting possibilities, limitations, and trade-offs. The approach, though realized as a specific Python implementation, is meant as a novel conceptual contribution, which can be partially or fully incorporated in other contexts. The library is available as open source at https://github.com/microsoft/tinytroupe.
comment: 9 pages. Preprint to be submitted to peer-review
☆ Te Ahorré Un Click: A Revised Definition of Clickbait and Detection in Spanish News
We revise the definition of clickbait, which lacks current consensus, and argue that the creation of a curiosity gap is the key concept that distinguishes clickbait from other related phenomena such as sensationalism and headlines that do not deliver what they promise or diverge from the article. Therefore, we propose a new definition: clickbait is a technique for generating headlines and teasers that deliberately omit part of the information with the goal of raising the readers' curiosity, capturing their attention and enticing them to click. We introduce a new approach to clickbait detection datasets creation, by refining the concept limits and annotations criteria, minimizing the subjectivity in the decision as much as possible. Following it, we created and release TA1C (for Te Ahorr\'e Un Click, Spanish for Saved You A Click), the first open source dataset for clickbait detection in Spanish. It consists of 3,500 tweets coming from 18 well known media sources, manually annotated and reaching a 0.825 Fleiss' K inter annotator agreement. We implement strong baselines that achieve 0.84 in F1-score.
☆ EventHunter: Dynamic Clustering and Ranking of Security Events from Hacker Forum Discussions RAID 2025
Hacker forums provide critical early warning signals for emerging cybersecurity threats, but extracting actionable intelligence from their unstructured and noisy content remains a significant challenge. This paper presents an unsupervised framework that automatically detects, clusters, and prioritizes security events discussed across hacker forum posts. Our approach leverages Transformer-based embeddings fine-tuned with contrastive learning to group related discussions into distinct security event clusters, identifying incidents like zero-day disclosures or malware releases without relying on predefined keywords. The framework incorporates a daily ranking mechanism that prioritizes identified events using quantifiable metrics reflecting timeliness, source credibility, information completeness, and relevance. Experimental evaluation on real-world hacker forum data demonstrates that our method effectively reduces noise and surfaces high-priority threats, enabling security analysts to mount proactive responses. By transforming disparate hacker forum discussions into structured, actionable intelligence, our work addresses fundamental challenges in automated threat detection and analysis.
comment: Accepted for publication at the 28th International Symposium on Research in Attacks, Intrusions, and Defenses (RAID 2025)
☆ Your Pretrained Model Tells the Difficulty Itself: A Self-Adaptive Curriculum Learning Paradigm for Natural Language Understanding ACL 2025
Curriculum learning is a widely adopted training strategy in natural language processing (NLP), where models are exposed to examples organized by increasing difficulty to enhance learning efficiency and performance. However, most existing approaches rely on manually defined difficulty metrics -- such as text length -- which may not accurately reflect the model's own perspective. To overcome this limitation, we present a self-adaptive curriculum learning paradigm that prioritizes fine-tuning examples based on difficulty scores predicted by pre-trained language models (PLMs) themselves. Building on these scores, we explore various training strategies that differ in the ordering of examples for the fine-tuning: from easy-to-hard, hard-to-easy, to mixed sampling. We evaluate our method on four natural language understanding (NLU) datasets covering both binary and multi-class classification tasks. Experimental results show that our approach leads to faster convergence and improved performance compared to standard random sampling.
comment: 18 pages, 23 figures. To appear in ACL 2025 Student Research Workshop (SRW)
☆ Sound and Complete Neuro-symbolic Reasoning with LLM-Grounded Interpretations
Large language models (LLMs) have demonstrated impressive capabilities in natural language understanding and generation, but they exhibit problems with logical consistency in the output they generate. How can we harness LLMs' broad-coverage parametric knowledge in formal reasoning despite their inconsistency? We present a method for directly integrating an LLM into the interpretation function of the formal semantics for a paraconsistent logic. We provide experimental evidence for the feasibility of the method by evaluating the function using datasets created from several short-form factuality benchmarks. Unlike prior work, our method offers a theoretical framework for neuro-symbolic reasoning that leverages an LLM's knowledge while preserving the underlying logic's soundness and completeness properties.
comment: 29 pages, 9 tables, 3 figures. Accepted to the 19th Conference on Neurosymbolic Learning and Reasoning (NeSy 2025)
☆ Large Language Models Encode Semantics in Low-Dimensional Linear Subspaces
Understanding the latent space geometry of large language models (LLMs) is key to interpreting their behavior and improving alignment. \baturay{However, it remains unclear to what extent LLMs internally organize representations related to semantic understanding. To investigate this, we conduct a large-scale empirical study of hidden states in transformer-based LLMs, analyzing 11 decoder-only models across 6 scientific topics and 12 layers each. We find that high-level semantic information consistently lies in low-dimensional subspaces that form linearly separable representations across distinct domains. This separability becomes more pronounced in deeper layers and under prompts that trigger structured reasoning or alignment behaviors$\unicode{x2013}$even when surface content is unchanged. This geometry enables simple yet effective causal interventions in hidden space; for example, reasoning patterns like chain-of-thought can be captured by a single vector direction. Together, these findings support the development of geometry-aware tools that operate directly on latent representations to detect and mitigate harmful or adversarial content, using methods such as transport-based defenses that leverage this separability. As a proof of concept, we demonstrate this potential by training a simple MLP classifier as a lightweight latent-space guardrail, which detects adversarial and malicious prompts with high precision.
☆ MCEval: A Dynamic Framework for Fair Multilingual Cultural Evaluation of LLMs
Large language models exhibit cultural biases and limited cross-cultural understanding capabilities, particularly when serving diverse global user populations. We propose MCEval, a novel multilingual evaluation framework that employs dynamic cultural question construction and enables causal analysis through Counterfactual Rephrasing and Confounder Rephrasing. Our comprehensive evaluation spans 13 cultures and 13 languages, systematically assessing both cultural awareness and cultural bias across different linguistic scenarios. The framework provides 39,897 cultural awareness instances and 17,940 cultural bias instances. Experimental results reveal performance disparities across different linguistic scenarios, demonstrating that optimal cultural performance is not only linked to training data distribution, but also is related to language-culture alignment. The evaluation results also expose the fairness issue, where approaches appearing successful in the English scenario create substantial disadvantages. MCEval represents the first comprehensive multilingual cultural evaluation framework that provides deeper insights into LLMs' cultural understanding.
☆ Towards Concise and Adaptive Thinking in Large Reasoning Models: A Survey
Large reasoning models (LRMs) like OpenAI o1 and DeepSeek R1 have demonstrated impressive performance on complex reasoning tasks like mathematics and programming with long Chain-of-Thought (CoT) reasoning sequences (slow-thinking), compared with traditional large language models (fast-thinking). However, these reasoning models also face a huge challenge that generating unnecessarily lengthy and redundant reasoning chains even for trivial questions. This phenomenon leads to a significant waste of inference resources, increases the response time for simple queries, and hinders the practical application of LRMs in real-world products. To this end, it is crucial to shorten lengthy reasoning chains and learn adaptive reasoning between fast and slow thinking based on input difficulty. In this survey, we provide a comprehensive overview of recent progress in concise and adaptive thinking for efficient reasoning of LRMs, including methodologies, benchmarks, and challenges for future exploration. We hope this survey can help researchers quickly understand the landscape of this field and inspire novel adaptive thinking ideas to facilitate better usage of LRMs.
☆ Can Group Relative Policy Optimization Improve Thai Legal Reasoning and Question Answering?
The Retrieval-Augmented Generation (RAG) systems' performance on Thai legal question answering is still limited, especially for questions requiring extensive, complex legal reasoning. To address these limitations, we introduce an approach aligning LLMs toward improved law citation accuracy and better response quality using Group-Relative Policy Optimization (GRPO). Our approach leverages BGE-M3 embeddings as a cost-efficient semantic-similarity reward, significantly reducing computational expenses up to 2.5x compared to large language model judges. Experiments on the NitiBench benchmark demonstrate substantial improvements: GRPO achieves up to 90% citation-F1 gains from the base model and a 31% increase in joint quality metrics over instruction tuning. Crucially, our method shows enhanced robustness on complex legal reasoning tasks compared to instruction tuning, providing an effective and resource-efficient solution for enhancing Thai legal LLMs.
☆ An Exploration of Knowledge Editing for Arabic
While Knowledge Editing (KE) has been widely explored in English, its behavior in morphologically rich languages like Arabic remains underexamined. In this work, we present the first study of Arabic KE. We evaluate four methods (ROME, MEMIT, ICE, and LTE) on Arabic translations of the ZsRE and Counterfact benchmarks, analyzing both multilingual and cross-lingual settings. Our experiments on Llama-2-7B-chat show show that parameter-based methods struggle with cross-lingual generalization, while instruction-tuned methods perform more robustly. We extend Learning-To-Edit (LTE) to a multilingual setting and show that joint Arabic-English training improves both editability and transfer. We release Arabic KE benchmarks and multilingual training for LTE data to support future research.
☆ SpreadPy: A Python tool for modelling spreading activation and superdiffusion in cognitive multiplex networks
We introduce SpreadPy as a Python library for simulating spreading activation in cognitive single-layer and multiplex networks. Our tool is designed to perform numerical simulations testing structure-function relationships in cognitive processes. By comparing simulation results with grounded theories in knowledge modelling, SpreadPy enables systematic investigations of how activation dynamics reflect cognitive, psychological and clinical phenomena. We demonstrate the library's utility through three case studies: (1) Spreading activation on associative knowledge networks distinguishes students with high versus low math anxiety, revealing anxiety-related structural differences in conceptual organization; (2) Simulations of a creativity task show that activation trajectories vary with task difficulty, exposing how cognitive load modulates lexical access; (3) In individuals with aphasia, simulated activation patterns on lexical networks correlate with empirical error types (semantic vs. phonological) during picture-naming tasks, linking network structure to clinical impairments. SpreadPy's flexible framework allows researchers to model these processes using empirically derived or theoretical networks, providing mechanistic insights into individual differences and cognitive impairments. The library is openly available, supporting reproducible research in psychology, neuroscience, and education research.
☆ NMIXX: Domain-Adapted Neural Embeddings for Cross-Lingual eXploration of Finance
General-purpose sentence embedding models often struggle to capture specialized financial semantics, especially in low-resource languages like Korean, due to domain-specific jargon, temporal meaning shifts, and misaligned bilingual vocabularies. To address these gaps, we introduce NMIXX (Neural eMbeddings for Cross-lingual eXploration of Finance), a suite of cross-lingual embedding models fine-tuned with 18.8K high-confidence triplets that pair in-domain paraphrases, hard negatives derived from a semantic-shift typology, and exact Korean-English translations. Concurrently, we release KorFinSTS, a 1,921-pair Korean financial STS benchmark spanning news, disclosures, research reports, and regulations, designed to expose nuances that general benchmarks miss. When evaluated against seven open-license baselines, NMIXX's multilingual bge-m3 variant achieves Spearman's rho gains of +0.10 on English FinSTS and +0.22 on KorFinSTS, outperforming its pre-adaptation checkpoint and surpassing other models by the largest margin, while revealing a modest trade-off in general STS performance. Our analysis further shows that models with richer Korean token coverage adapt more effectively, underscoring the importance of tokenizer design in low-resource, cross-lingual settings. By making both models and the benchmark publicly available, we provide the community with robust tools for domain-adapted, multilingual representation learning in finance.
comment: Under Review
☆ MENTOR: Efficient Multimodal-Conditioned Tuning for Autoregressive Vision Generation Models
Recent text-to-image models produce high-quality results but still struggle with precise visual control, balancing multimodal inputs, and requiring extensive training for complex multimodal image generation. To address these limitations, we propose MENTOR, a novel autoregressive (AR) framework for efficient Multimodal-conditioned Tuning for Autoregressive multimodal image generation. MENTOR combines an AR image generator with a two-stage training paradigm, enabling fine-grained, token-level alignment between multimodal inputs and image outputs without relying on auxiliary adapters or cross-attention modules. The two-stage training consists of: (1) a multimodal alignment stage that establishes robust pixel- and semantic-level alignment, followed by (2) a multimodal instruction tuning stage that balances the integration of multimodal inputs and enhances generation controllability. Despite modest model size, suboptimal base components, and limited training resources, MENTOR achieves strong performance on the DreamBench++ benchmark, outperforming competitive baselines in concept preservation and prompt following. Additionally, our method delivers superior image reconstruction fidelity, broad task adaptability, and improved training efficiency compared to diffusion-based methods. Dataset, code, and models are available at: https://github.com/HaozheZhao/MENTOR
comment: 24 pages,12 figures
☆ Adapting Definition Modeling for New Languages: A Case Study on Belarusian
Definition modeling, the task of generating new definitions for words in context, holds great prospect as a means to assist the work of lexicographers in documenting a broader variety of lects and languages, yet much remains to be done in order to assess how we can leverage pre-existing models for as-of-yet unsupported languages. In this work, we focus on adapting existing models to Belarusian, for which we propose a novel dataset of 43,150 definitions. Our experiments demonstrate that adapting a definition modeling systems requires minimal amounts of data, but that there currently are gaps in what automatic metrics do capture.
comment: To appear at SlavicNLP 2025
☆ How Important is `Perfect' English for Machine Translation Prompts?
Large language models (LLMs) have achieved top results in recent machine translation evaluations, but they are also known to be sensitive to errors and perturbations in their prompts. We systematically evaluate how both humanly plausible and synthetic errors in user prompts affect LLMs' performance on two related tasks: Machine translation and machine translation evaluation. We provide both a quantitative analysis and qualitative insights into how the models respond to increasing noise in the user prompt. The prompt quality strongly affects the translation performance: With many errors, even a good prompt can underperform a minimal or poor prompt without errors. However, different noise types impact translation quality differently, with character-level and combined noisers degrading performance more than phrasal perturbations. Qualitative analysis reveals that lower prompt quality largely leads to poorer instruction following, rather than directly affecting translation quality itself. Further, LLMs can still translate in scenarios with overwhelming random noise that would make the prompt illegible to humans.
☆ Ref-Long: Benchmarking the Long-context Referencing Capability of Long-context Language Models ACL 2025
Long-context language models (LCLMs) have exhibited impressive capabilities in long-context understanding tasks. Among these, long-context referencing -- a crucial task that requires LCLMs to attribute items of interest to specific parts of long-context data -- remains underexplored. To bridge this gap, this paper proposes Referencing Evaluation for Long-context Language Models (Ref-Long), a novel benchmark designed to assess the long-context referencing capability of LCLMs. Specifically, Ref-Long requires LCLMs to identify the indexes of documents that reference a specific key, emphasizing contextual relationships between the key and the documents over simple retrieval. Based on the task design, we construct three subsets ranging from synthetic to realistic scenarios to form the Ref-Long benchmark. Experimental results of 13 LCLMs reveal significant shortcomings in long-context referencing, even among advanced models like GPT-4o. To further investigate these challenges, we conduct comprehensive analyses, including human evaluations, task format adjustments, fine-tuning experiments, and error analyses, leading to several key insights. Our data and code can be found in https://github. com/wujunjie1998/Ref-Long.
comment: ACL 2025 Main Conference. First 2 authors contributed equally
☆ GoalfyMax: A Protocol-Driven Multi-Agent System for Intelligent Experience Entities
Modern enterprise environments demand intelligent systems capable of handling complex, dynamic, and multi-faceted tasks with high levels of autonomy and adaptability. However, traditional single-purpose AI systems often lack sufficient coordination, memory reuse, and task decomposition capabilities, limiting their scalability in realistic settings. To address these challenges, we present \textbf{GoalfyMax}, a protocol-driven framework for end-to-end multi-agent collaboration. GoalfyMax introduces a standardized Agent-to-Agent (A2A) communication layer built on the Model Context Protocol (MCP), allowing independent agents to coordinate through asynchronous, protocol-compliant interactions. It incorporates the Experience Pack (XP) architecture, a layered memory system that preserves both task rationales and execution traces, enabling structured knowledge retention and continual learning. Moreover, our system integrates advanced features including multi-turn contextual dialogue, long-short term memory modules, and dynamic safety validation, supporting robust, real-time strategy adaptation. Empirical results on complex task orchestration benchmarks and case study demonstrate that GoalfyMax achieves superior adaptability, coordination, and experience reuse compared to baseline frameworks. These findings highlight its potential as a scalable, future-ready foundation for multi-agent intelligent systems.
☆ Balanced Training Data Augmentation for Aspect-Based Sentiment Analysis
Aspect-based sentiment analysis (ABSA) is a crucial fine-grained task in social media scenarios to identify the sentiment polarity of specific aspect terms in a sentence. Although many existing studies leverage large language models (LLMs) to perform ABSA due to their strong context understanding capabilities, they still face challenges to learn the context information in the running text because of the short text, as well as the small and unbalanced labeled training data, where most data are labeled with positive sentiment. Data augmentation (DA) is a feasible strategy for providing richer contextual information, especially when using LLMs to create synthetic training data, but faces challenges in ensuring a high quality of the augmented data.In this paper, we propose an LLM-based ABSA approach with training data augmentation.Specifically, an LLM is prompted to generate augmented training data based on the original training data, so as to construct a new training data with larger size and balanced label distributions to better train an ABSA model. Meanwhile, in order to improve the quality of the augmented data, we propose a reinforcement learning approach to optimize the data augmentation. LLM.Experiment results and further analyses on English benchmark datasets for ABSA demonstrate the effectiveness of our approach, where superior performance is observed over strong baselines and most existing studies.
☆ ViSP: A PPO-Driven Framework for Sarcasm Generation with Contrastive Learning
Human emotions are complex, with sarcasm being a subtle and distinctive form. Despite progress in sarcasm research, sarcasm generation remains underexplored, primarily due to the overreliance on textual modalities and the neglect of visual cues, as well as the mismatch between image content and sarcastic intent in existing datasets. In this paper, we introduce M2SaG, a multimodal sarcasm generation dataset with 4,970 samples, each containing an image, a sarcastic text, and a sarcasm target. To benchmark M2SaG, we propose ViSP, a generation framework that integrates Proximal Policy Optimization (PPO) and contrastive learning. PPO utilizes reward scores from DIP to steer the generation of sarcastic texts, while contrastive learning encourages the model to favor outputs with higher reward scores. These strategies improve overall generation quality and produce texts with more pronounced sarcastic intent. We evaluate ViSP across five metric sets and find it surpasses all baselines, including large language models, underscoring their limitations in sarcasm generation. Furthermore, we analyze the distributions of Sarcasm Scores and Factual Incongruity for both M2SaG and the texts generated by ViSP. The generated texts exhibit higher mean Sarcasm Scores (0.898 vs. 0.770) and Factual Incongruity (0.768 vs. 0.739), demonstrating that ViSP produces higher-quality sarcastic content than the original dataset. % The dataset and code will be publicly available. Our dataset and code will be released at \textit{https://github.com/wclapply/ViSP}.
☆ Evaluating LLMs on Sequential API Call Through Automated Test Generation
By integrating tools from external APIs, Large Language Models (LLMs) have expanded their promising capabilities in a diverse spectrum of complex real-world tasks. However, testing, evaluation, and analysis of LLM tool use remain in their early stages. Most existing benchmarks rely on manually collected test cases, many of which cannot be automatically checked for semantic correctness and instead depend on static methods such as string matching. Additionally, these benchmarks often overlook the complex interactions that occur between sequential API calls, which are common in real-world applications. To fill the gap, in this paper, we introduce StateGen, an automated framework designed to generate diverse coding tasks involving sequential API interactions. StateGen combines state-machine-based API constraint solving and validation, energy-based sampling, and control-flow injection to generate executable programs. These programs are then translated into human-like natural language task descriptions through a collaboration of two LLM agents. Utilizing StateGen, we construct StateEval, a benchmark encompassing 120 verified test cases spanning across three representative scenarios: Session Service, Tensor Operation, and ElevenLabs MCP. Experimental results confirm that StateGen can effectively generate challenging and realistic API-oriented tasks, highlighting areas for improvement in current LLMs incorporating APIs.
☆ Towards Agentic RAG with Deep Reasoning: A Survey of RAG-Reasoning Systems in LLMs
Retrieval-Augmented Generation (RAG) lifts the factuality of Large Language Models (LLMs) by injecting external knowledge, yet it falls short on problems that demand multi-step inference; conversely, purely reasoning-oriented approaches often hallucinate or mis-ground facts. This survey synthesizes both strands under a unified reasoning-retrieval perspective. We first map how advanced reasoning optimizes each stage of RAG (Reasoning-Enhanced RAG). Then, we show how retrieved knowledge of different type supply missing premises and expand context for complex inference (RAG-Enhanced Reasoning). Finally, we spotlight emerging Synergized RAG-Reasoning frameworks, where (agentic) LLMs iteratively interleave search and reasoning to achieve state-of-the-art performance across knowledge-intensive benchmarks. We categorize methods, datasets, and open challenges, and outline research avenues toward deeper RAG-Reasoning systems that are more effective, multimodally-adaptive, trustworthy, and human-centric. The collection is available at https://github.com/DavidZWZ/Awesome-RAG-Reasoning.
comment: submitted to ARR May
☆ The CoNLL-2013 Shared Task on Grammatical Error Correction
The CoNLL-2013 shared task was devoted to grammatical error correction. In this paper, we give the task definition, present the data sets, and describe the evaluation metric and scorer used in the shared task. We also give an overview of the various approaches adopted by the participating teams, and present the evaluation results.
comment: 12 pages
☆ Enhancing Clinical Text Classification via Fine-Tuned DRAGON Longformer Models
This study explores the optimization of the DRAGON Longformer base model for clinical text classification, specifically targeting the binary classification of medical case descriptions. A dataset of 500 clinical cases containing structured medical observations was used, with 400 cases for training and 100 for validation. Enhancements to the pre-trained joeranbosma/dragon-longformer-base-mixed-domain model included hyperparameter tuning, domain-specific preprocessing, and architectural adjustments. Key modifications involved increasing sequence length from 512 to 1024 tokens, adjusting learning rates from 1e-05 to 5e-06, extending training epochs from 5 to 8, and incorporating specialized medical terminology. The optimized model achieved notable performance gains: accuracy improved from 72.0% to 85.2%, precision from 68.0% to 84.1%, recall from 75.0% to 86.3%, and F1-score from 71.0% to 85.2%. Statistical analysis confirmed the significance of these improvements (p < .001). The model demonstrated enhanced capability in interpreting medical terminology, anatomical measurements, and clinical observations. These findings contribute to domain-specific language model research and offer practical implications for clinical natural language processing applications. The optimized model's strong performance across diverse medical conditions underscores its potential for broad use in healthcare settings.
comment: 29 pages, 5 tables
♻ ☆ Beyond Multiple Choice: Evaluating Steering Vectors for Adaptive Free-Form Summarization ICML 2025
Steering vectors are a lightweight method for controlling text properties by adding a learned bias to language model activations at inference time. So far, steering vectors have predominantly been evaluated in multiple-choice settings, while their effectiveness in free-form generation tasks remains understudied. Moving "Beyond Multiple Choice," we thoroughly evaluate the effectiveness of steering vectors in adaptively controlling topical focus, sentiment, toxicity, and readability in abstractive summaries of the NEWTS dataset. We find that steering effectively controls the targeted summary properties, but high steering strengths consistently degrade both intrinsic and extrinsic text quality. Compared to steering, prompting offers weaker control, while preserving text quality. Combining steering and prompting yields the strongest control over text properties and offers the most favorable efficacy-quality trade-off at moderate steering strengths. Our results underscore the practical trade-off between control strength and text quality preservation when applying steering vectors to free-form generation tasks.
comment: 29 pages, 21 figures, published at ICML 2025 Workshop on Reliable and Responsible Foundation Models
♻ ☆ VisOnlyQA: Large Vision Language Models Still Struggle with Visual Perception of Geometric Information
Large Vision Language Models (LVLMs) have achieved remarkable performance in various vision-language tasks. However, it is still unclear how accurately LVLMs can perceive visual information in images. In particular, the capability of LVLMs to perceive geometric information, such as shape, angle, and size, remains insufficiently analyzed, although the perception of these properties is crucial for tasks that require a detailed visual understanding. In this work, we introduce VisOnlyQA, a dataset for evaluating the geometric perception of LVLMs, and reveal that LVLMs often cannot accurately perceive basic geometric information in images, while human performance is nearly perfect. VisOnlyQA consists of 12 tasks that directly ask about geometric information in geometric shapes, charts, chemical structures, and 3D shapes. Our experiments highlight the following findings: (i) State-of-the-art LVLMs struggle with basic geometric perception. 23 LVLMs we evaluate, including GPT-4o and Gemini 2.5 Pro, work poorly on VisOnlyQA. (ii) Additional training data does not resolve this issue. Fine-tuning on the training set of VisOnlyQA is not always effective, even for in-distribution tasks. (iii) LLM may be the bottleneck. LVLMs using stronger LLMs exhibit better geometric perception on VisOnlyQA, while it does not require complex reasoning, suggesting that the way LVLMs process information from visual encoders is a bottleneck. The datasets, code, and model responses are provided at https://github.com/psunlpgroup/VisOnlyQA.
comment: COLM 2025. VisOnlyQA dataset, code, and model responses are provided at https://github.com/psunlpgroup/VisOnlyQA. Please also refer to our project website at https://visonlyqa.github.io/
♻ ☆ SymbolicThought: Integrating Language Models and Symbolic Reasoning for Consistent and Interpretable Human Relationship Understanding
Understanding character relationships is essential for interpreting complex narratives and conducting socially grounded AI research. However, manual annotation is time-consuming and low in coverage, while large language models (LLMs) often produce hallucinated or logically inconsistent outputs. We present SymbolicThought, a human-in-the-loop framework that combines LLM-based extraction with symbolic reasoning. The system constructs editable character relationship graphs, refines them using seven types of logical constraints, and enables real-time validation and conflict resolution through an interactive interface. To support logical supervision and explainable social analysis, we release a dataset of 160 interpersonal relationships with corresponding logical structures. Experiments show that SymbolicThought improves annotation accuracy and consistency while significantly reducing time cost, offering a practical tool for narrative understanding, explainable AI, and LLM evaluation.
♻ ☆ LASER: Attention with Exponential Transformation ICML 2025
Transformers have had tremendous impact for several sequence related tasks, largely due to their ability to retrieve from any part of the sequence via softmax based dot-product attention. This mechanism plays a crucial role in Transformer's performance. We analyze the gradients backpropagated through the softmax operation in the attention mechanism and observe that these gradients can often be small. This poor gradient signal backpropagation can lead to inefficient learning of parameters preceeding the attention operations. To this end, we introduce a new attention mechanism called LASER, which we analytically show to admit a larger gradient signal. We show that LASER attention can be implemented by making small modifications to existing attention implementations. We conduct experiments on autoregressive large language models (LLMs) with upto 7.7 billion parameters with an average improvement of upto 1.44% over standard attention on downstream evaluations and 1.65% finetuning improvements. Additionally, LASER demonstrates generalization performance improvement across a variety of tasks (vision, text and speech):Vision Transformer (ViT) on Imagenet, Conformer on the Librispeech speech-to-text and BERT with 2.2 billion parameters.
comment: ICML 2025
♻ ☆ DataDecide: How to Predict Best Pretraining Data with Small Experiments ICML 2025
Because large language models are expensive to pretrain on different datasets, using smaller-scale experiments to decide on data is crucial for reducing costs. Which benchmarks and methods of making decisions from observed performance at small scale most accurately predict the datasets that yield the best large models? To empower open exploration of this question, we release models, data, and evaluations in DataDecide -- the most extensive open suite of models over differences in data and scale. We conduct controlled pretraining experiments across 25 corpora with differing sources, deduplication, and filtering up to 100B tokens, model sizes up to 1B parameters, and 3 random seeds. We find that the ranking of models at a single, small size (e.g., 150M parameters) is a strong baseline for predicting best models at our larger target scale (1B) (~80% of com parisons correct). No scaling law methods among 8 baselines exceed the compute-decision frontier of single-scale predictions, but DataDecide can measure improvement in future scaling laws. We also identify that using continuous likelihood metrics as proxies in small experiments makes benchmarks including MMLU, ARC, HellaSwag, MBPP, and HumanEval >80% predictable at the target 1B scale with just 0.01% of the compute.
comment: ICML 2025
♻ ☆ Cascade Speculative Drafting for Even Faster LLM Inference NeurIPS 2024
Introduced to enhance the efficiency of large language model (LLM) inference, speculative decoding operates by having a smaller model generate a draft. A larger target model then reviews this draft to align with its output, and any acceptance by the target model results in a reduction of the number of the target model runs, ultimately improving efficiency. However, the drafting process in speculative decoding includes slow autoregressive generation and allocates equal time to generating tokens, irrespective of their importance. These inefficiencies collectively contribute to the suboptimal performance of speculative decoding. To further improve LLM inference, we introduce Cascade Speculative Drafting (CS Drafting), a speculative execution algorithm that incorporates two types of cascades. The Vertical Cascade eliminates autoregressive generation from neural models, while the Horizontal Cascade optimizes time allocation in drafting for improved efficiency. Combining both cascades, CS Drafting achieves greater speedup compared to the baselines in our experiments, while preserving the same output distribution as the target model.
comment: NeurIPS 2024
♻ ☆ KnowShiftQA: How Robust are RAG Systems when Textbook Knowledge Shifts in K-12 Education? ACL 2025
Retrieval-Augmented Generation (RAG) systems show remarkable potential as question answering tools in the K-12 Education domain, where knowledge is typically queried within the restricted scope of authoritative textbooks. However, discrepancies between these textbooks and the parametric knowledge inherent in Large Language Models (LLMs) can undermine the effectiveness of RAG systems. To systematically investigate RAG system robustness against such knowledge discrepancies, we introduce KnowShiftQA. This novel question answering dataset simulates these discrepancies by applying deliberate hypothetical knowledge updates to both answers and source documents, reflecting how textbook knowledge can shift. KnowShiftQA comprises 3,005 questions across five subjects, designed with a comprehensive question typology focusing on context utilization and knowledge integration. Our extensive experiments on retrieval and question answering performance reveal that most RAG systems suffer a substantial performance drop when faced with these knowledge discrepancies. Furthermore, questions requiring the integration of contextual (textbook) knowledge with parametric (LLM) knowledge pose a significant challenge to current LLMs.
comment: ACL 2025 Main
♻ ☆ From Fragments to Facts: A Curriculum-Driven DPO Approach for Generating Hindi News Veracity Explanations
In an era of rampant misinformation, generating reliable news explanations is vital, especially for under-represented languages like Hindi. Lacking robust automated tools, Hindi faces challenges in scaling misinformation detection. To bridge this gap, we propose a novel framework integrating Direct Preference Optimization (DPO) with curriculum learning to align machine-generated explanations with human reasoning. Fact-checked explanations from credible sources serve as preferred responses, while LLM outputs highlight system limitations and serve as non-preferred responses. To refine task-specific alignment, we introduce two key parameters -- Actuality and Finesse -- into the DPO loss function, enhancing explanation quality and consistency. Experiments with LLMs (Mistral, Llama, Gemma) and PLMs (mBART, mT5) confirm the framework's effectiveness in generating coherent, contextually relevant explanations. This scalable approach combats misinformation and extends automated explanation generation to low-resource languages.
♻ ☆ Reinforcing Question Answering Agents with Minimalist Policy Gradient Optimization
Large Language Models (LLMs) have demonstrated remarkable versatility, due to the lack of factual knowledge, their application to Question Answering (QA) tasks remains hindered by hallucination. While Retrieval-Augmented Generation mitigates these issues by integrating external knowledge, existing approaches rely heavily on in-context learning, whose performance is constrained by the fundamental reasoning capabilities of LLMs. In this paper, we propose Mujica, a Multi-hop Joint Intelligence for Complex Question Answering, comprising a planner that decomposes questions into a directed acyclic graph of subquestions and a worker that resolves questions via retrieval and reasoning. Additionally, we introduce MyGO (Minimalist policy Gradient Optimization), a novel reinforcement learning method that replaces traditional policy gradient updates with Maximum Likelihood Estimation (MLE) by sampling trajectories from an asymptotically optimal policy. MyGO eliminates the need for gradient rescaling and reference models, ensuring stable and efficient training. Empirical results across multiple datasets demonstrate the effectiveness of Mujica-MyGO in enhancing multi-hop QA performance for various LLMs, offering a scalable and resource-efficient solution for complex QA tasks.
♻ ☆ Perception-Aware Policy Optimization for Multimodal Reasoning
Reinforcement Learning with Verifiable Rewards (RLVR) has proven to be a highly effective strategy for endowing Large Language Models (LLMs) with robust multi-step reasoning abilities. However, its design and optimizations remain tailored to purely textual domains, resulting in suboptimal performance when applied to multimodal reasoning tasks. In particular, we observe that a major source of error in current multimodal reasoning lies in the perception of visual inputs. To address this bottleneck, we propose Perception-Aware Policy Optimization (PAPO), a simple yet effective extension of GRPO that encourages the model to learn to perceive while learning to reason, entirely from internal supervision signals. Notably, PAPO does not rely on additional data curation, external reward models, or proprietary models. Specifically, we introduce the Implicit Perception Loss in the form of a KL divergence term to the GRPO objective, which, despite its simplicity, yields significant overall improvements (4.4%) on diverse multimodal benchmarks. The improvements are more pronounced, approaching 8.0%, on tasks with high vision dependency. We also observe a substantial reduction (30.5%) in perception errors, indicating improved perceptual capabilities with PAPO. We conduct comprehensive analysis of PAPO and identify a unique loss hacking issue, which we rigorously analyze and mitigate through a Double Entropy Loss. Overall, our work introduces a deeper integration of perception-aware supervision into RLVR learning objectives and lays the groundwork for a new RL framework that encourages visually grounded reasoning. Project page: https://mikewangwzhl.github.io/PAPO.
♻ ☆ Teaching Models to Verbalize Reward Hacking in Chain-of-Thought Reasoning ICML 2025
Language models trained with reinforcement learning (RL) can engage in reward hacking--the exploitation of unintended strategies for high reward--without revealing this behavior in their chain-of-thought reasoning. This makes the detection of reward hacking difficult, posing risks for high-stakes applications. We propose verbalization fine-tuning (VFT), a pre-RL fine-tuning intervention that trains models to explicitly acknowledge when they are influenced by prompt cues--hints which point to incorrect answers (e.g., "a Stanford professor thinks the answer is A"). To evaluate VFT, we subsequently train models with RL on environments where held-out prompt cues signal which incorrect answers will receive high reward, incentivizing models to exploit these cues instead of reasoning correctly. We measure how often models exploit these cues without verbalizing it. After RL, only 6% of the VFT-trained model's responses consist of undetected reward hacks. In comparison, when we perform RL without VFT, the rate of undetected reward hacks goes up to 88%; with a debiasing baseline intervention, this increases further to 99%. VFT achieves this by substantially increasing how often models verbalize the influence of cues, from 8% to 43% after VFT, and up to 94% after RL. Baselines remain low even after RL (11% and 1%). Our results show that teaching models to explicitly verbalize reward hacking behavior before RL significantly improves their detection, offering a practical path toward more transparent and safe AI systems.
comment: Published at ICML 2025 Workshop on Reliable and Responsible Foundation Models
♻ ☆ Learning-to-Context Slope: Evaluating In-Context Learning Effectiveness Beyond Performance Illusions
In-context learning (ICL) has emerged as an effective approach to enhance the performance of large language models (LLMs). However, its effectiveness varies significantly across models and tasks, posing challenges for practitioners to determine when ICL reliably improves performance. Current evaluation approaches, reliant on performance change after applying ICL, suffer from low reliability, poor attribution, and impracticality in data-insufficient scenarios. We propose the Learning-to-Context Slope (LCS), a novel metric that quantifies ICL effectiveness by modeling the slope between learning gain (loss decrease from demonstrations) and contextual relevance (demonstration-input relevance). LCS addresses key limitations of performance-based metrics: (1) it captures continuous loss changes even when outputs are incorrect, improving reliability; (2) its formulation attributes ICL failures to weak contextual alignment (inability to adapt inputs to demonstrations) or strong output calibration (self-verification of correctness); and (3) it minimizes reliance on labeled data via synthetic evaluation. Extensive experiments demonstrate that LCS strongly correlates with performance improvements in labeled settings and reliably reflects true effectiveness in biased or data-scarce scenarios. Further analysis reveals actionable thresholds for LCS and identifies model capabilities critical to ICL success.
♻ ☆ OmniSQL: Synthesizing High-quality Text-to-SQL Data at Scale
Text-to-SQL, the task of translating natural language questions into SQL queries, plays a crucial role in enabling non-experts to interact with databases. While recent advancements in large language models (LLMs) have significantly enhanced text-to-SQL performance, existing approaches face notable limitations in real-world text-to-SQL applications. Prompting-based methods often depend on closed-source LLMs, which are expensive, raise privacy concerns, and lack customization. Fine-tuning-based methods, on the other hand, suffer from poor generalizability due to the limited coverage of publicly available training data. To overcome these challenges, we propose a novel and scalable text-to-SQL data synthesis framework for automatically synthesizing large-scale, high-quality, and diverse datasets without extensive human intervention. Using this framework, we introduce SynSQL-2.5M, the first million-scale text-to-SQL dataset, containing 2.5 million samples spanning over 16,000 synthetic databases. Each sample includes a database, SQL query, natural language question, and chain-of-thought (CoT) solution. Leveraging SynSQL-2.5M, we develop OmniSQL, a powerful open-source text-to-SQL model available in three sizes: 7B, 14B, and 32B. Extensive evaluations across nine datasets demonstrate that OmniSQL achieves state-of-the-art performance, matching or surpassing leading closed-source and open-source LLMs, including GPT-4o and DeepSeek-V3, despite its smaller size. We release all code, datasets, and models to support further research.
♻ ☆ MoRE: A Mixture of Reflectors Framework for Large Language Model-Based Sequential Recommendation RecSys'25
Large language models (LLMs) have emerged as a cutting-edge approach in sequential recommendation, leveraging historical interactions to model dynamic user preferences. Current methods mainly focus on learning processed recommendation data in the form of sequence-to-sequence text. While effective, they exhibit three key limitations: 1) failing to decouple intra-user explicit features (e.g., product titles) from implicit behavioral patterns (e.g., brand loyalty) within interaction histories; 2) underutilizing cross-user collaborative filtering (CF) signals; and 3) relying on inefficient reflection update strategies. To address this, We propose MoRE (Mixture of REflectors), which introduces three perspective-aware offline reflection processes to address these gaps. This decomposition directly resolves Challenges 1 (explicit/implicit ambiguity) and 2 (CF underutilization). Furthermore, MoRE's meta-reflector employs a self-improving strategy and a dynamic selection mechanism (Challenge 3) to adapt to evolving user preferences. First, two intra-user reflectors decouple explicit and implicit patterns from a user's interaction sequence, mimicking traditional recommender systems' ability to distinguish surface-level and latent preferences. A third cross-user reflector captures CF signals by analyzing user similarity patterns from multiple users' interactions. To optimize reflection quality, MoRE's meta-reflector employs a offline self-improving strategy that evaluates reflection impacts through comparisons of presence/absence and iterative refinement of old/new versions, with a online contextual bandit mechanism dynamically selecting the optimal perspective for recommendation for each user. Code: https://github.com/E-qin/MoRE-Rec.
comment: First 2 authors contributes equally to this work, accepted by RecSys'25 spotlight oral. Corresponding author is Weijie Yu(yu@uibe.edu.cn)
♻ ☆ Your Absorbing Discrete Diffusion Secretly Models the Bayesian Posterior
Discrete diffusion language models learn to reconstruct text from randomly masked inputs, yet under mild assumptions their denoiser already implements the exact Bayesian posterior over the original tokens. We prove that the expected denoiser output under the forward corruption distribution recovers the true posterior, and that a simple Monte Carlo estimator converges to this posterior at rate O(1/sqrt(K)) with finite-sample concentration bounds. Building on this insight, we introduce an inference-time ensemble that runs K independent denoising passes and aggregates both posterior means and variances without any extra training. On WikiText-2, our MC-marginal sampler recovers the analytic lambda-DCE zero-shot perplexity (approximately 39) to within a few points at K=128, and its per-token variance shows a strong rank correlation with reconstruction error (Spearman rho = 0.996). This cost-proportional procedure yields calibrated uncertainty estimates and a direct trade-off between compute and posterior fidelity in discrete diffusion LMs.
comment: 12 pages, 2 figures, 2 tables
♻ ☆ Knowledge-Augmented Multimodal Clinical Rationale Generation for Disease Diagnosis with Small Language Models
Interpretation is critical for disease diagnosis, but existing models struggle to balance predictive accuracy with human-understandable rationales. While large language models (LLMs) offer strong reasoning abilities, their clinical use is limited by high computational costs and restricted multimodal reasoning ability. Small language models (SLMs) are efficient but lack advanced reasoning for integrating multimodal medical data. In addition, both LLMs and SLMs lack domain knowledge for trustworthy reasoning. Therefore, we propose ClinRaGen, enhancing SLMs by leveraging LLM-derived reasoning ability via rationale distillation and domain knowledge injection for trustworthy multimodal rationale generation. Key innovations include a sequential rationale distillation framework that equips SLMs with LLM-comparable multimodal reasoning abilities, and a knowledge-augmented attention mechanism that jointly unifies multimodal representation from time series and textual data in the same encoding space, enabling it to be naturally interpreted by SLMs while incorporating domain knowledge for reliable rationale generation. Experiments on real-world medical datasets show that ClinRaGen achieves state-of-the-art performance in disease diagnosis and rationale generation, demonstrating the effectiveness of combining LLM-driven reasoning with knowledge augmentation for improved interpretability.
comment: 13 pages. 7 figures
♻ ☆ Large Language Model Psychometrics: A Systematic Review of Evaluation, Validation, and Enhancement
The advancement of large language models (LLMs) has outpaced traditional evaluation methodologies. This progress presents novel challenges, such as measuring human-like psychological constructs, moving beyond static and task-specific benchmarks, and establishing human-centered evaluation. These challenges intersect with psychometrics, the science of quantifying the intangible aspects of human psychology, such as personality, values, and intelligence. This review paper introduces and synthesizes the emerging interdisciplinary field of LLM Psychometrics, which leverages psychometric instruments, theories, and principles to evaluate, understand, and enhance LLMs. The reviewed literature systematically shapes benchmarking principles, broadens evaluation scopes, refines methodologies, validates results, and advances LLM capabilities. Diverse perspectives are integrated to provide a structured framework for researchers across disciplines, enabling a more comprehensive understanding of this nascent field. Ultimately, the review provides actionable insights for developing future evaluation paradigms that align with human-level AI and promote the advancement of human-centered AI systems for societal benefit. A curated repository of LLM psychometric resources is available at https://github.com/valuebyte-ai/Awesome-LLM-Psychometrics.
comment: 474 references
♻ ☆ Can A Society of Generative Agents Simulate Human Behavior and Inform Public Health Policy? A Case Study on Vaccine Hesitancy
Can we simulate a sandbox society with generative agents to model human behavior, thereby reducing the over-reliance on real human trials for assessing public policies? In this work, we investigate the feasibility of simulating health-related decision-making, using vaccine hesitancy, defined as the delay in acceptance or refusal of vaccines despite the availability of vaccination services (MacDonald, 2015), as a case study. To this end, we introduce the VacSim framework with 100 generative agents powered by Large Language Models (LLMs). VacSim simulates vaccine policy outcomes with the following steps: 1) instantiate a population of agents with demographics based on census data; 2) connect the agents via a social network and model vaccine attitudes as a function of social dynamics and disease-related information; 3) design and evaluate various public health interventions aimed at mitigating vaccine hesitancy. To align with real-world results, we also introduce simulation warmup and attitude modulation to adjust agents' attitudes. We propose a series of evaluations to assess the reliability of various LLM simulations. Experiments indicate that models like Llama and Qwen can simulate aspects of human behavior but also highlight real-world alignment challenges, such as inconsistent responses with demographic profiles. This early exploration of LLM-driven simulations is not meant to serve as definitive policy guidance; instead, it serves as a call for action to examine social simulation for policy development.
comment: Accepted to COLM 2025
♻ ☆ READoc: A Unified Benchmark for Realistic Document Structured Extraction ACL 2025
Document Structured Extraction (DSE) aims to extract structured content from raw documents. Despite the emergence of numerous DSE systems, their unified evaluation remains inadequate, significantly hindering the field's advancement. This problem is largely attributed to existing benchmark paradigms, which exhibit fragmented and localized characteristics. To address these limitations and offer a thorough evaluation of DSE systems, we introduce a novel benchmark named READoc, which defines DSE as a realistic task of converting unstructured PDFs into semantically rich Markdown. The READoc dataset is derived from 3,576 diverse and real-world documents from arXiv, GitHub, and Zenodo. In addition, we develop a DSE Evaluation S$^3$uite comprising Standardization, Segmentation and Scoring modules, to conduct a unified evaluation of state-of-the-art DSE approaches. By evaluating a range of pipeline tools, expert visual models, and general VLMs, we identify the gap between current work and the unified, realistic DSE objective for the first time. We aspire that READoc will catalyze future research in DSE, fostering more comprehensive and practical solutions.
comment: ACL 2025 Findings
♻ ☆ IDEAL: Influence-Driven Selective Annotations Empower In-Context Learners in Large Language Models ICLR 2024
In-context learning is a promising paradigm that utilizes in-context examples as prompts for the predictions of large language models. These prompts are crucial for achieving strong performance. However, since the prompts need to be sampled from a large volume of annotated examples, finding the right prompt may result in high annotation costs. To address this challenge, this paper introduces an influence-driven selective annotation method that aims to minimize annotation costs while improving the quality of in-context examples. The essence of our method is to select a pivotal subset from a large-scale unlabeled data pool to annotate for the subsequent sampling of prompts. Specifically, a directed graph is first constructed to represent unlabeled data. Afterward, the influence of candidate unlabeled subsets is quantified with a diffusion process. A simple yet effective greedy algorithm for unlabeled data selection is lastly introduced. It iteratively selects the data if it provides a maximum marginal gain with respect to quantified influence. Compared with previous efforts on selective annotations, our influence-driven method works in an end-to-end manner, avoids an intractable explicit balance between data diversity and representativeness, and enjoys theoretical support. Experiments confirm the superiority of the proposed method on various benchmarks, achieving better performance under lower time consumption during subset selection. The project page is available at https://skzhang1.github.io/IDEAL/.
comment: Accepted by ICLR 2024
♻ ☆ Topic Modeling as Multi-Objective Contrastive Optimization ICLR 2024
Recent representation learning approaches enhance neural topic models by optimizing the weighted linear combination of the evidence lower bound (ELBO) of the log-likelihood and the contrastive learning objective that contrasts pairs of input documents. However, document-level contrastive learning might capture low-level mutual information, such as word ratio, which disturbs topic modeling. Moreover, there is a potential conflict between the ELBO loss that memorizes input details for better reconstruction quality, and the contrastive loss which attempts to learn topic representations that generalize among input documents. To address these issues, we first introduce a novel contrastive learning method oriented towards sets of topic vectors to capture useful semantics that are shared among a set of input documents. Secondly, we explicitly cast contrastive topic modeling as a gradient-based multi-objective optimization problem, with the goal of achieving a Pareto stationary solution that balances the trade-off between the ELBO and the contrastive objective. Extensive experiments demonstrate that our framework consistently produces higher-performing neural topic models in terms of topic coherence, topic diversity, and downstream performance.
comment: Accepted at ICLR 2024 (poster). Official version available at: https://openreview.net/forum?id=HdAoLSBYXj
♻ ☆ Auditing Prompt Caching in Language Model APIs ICML 2025
Prompt caching in large language models (LLMs) results in data-dependent timing variations: cached prompts are processed faster than non-cached prompts. These timing differences introduce the risk of side-channel timing attacks. For example, if the cache is shared across users, an attacker could identify cached prompts from fast API response times to learn information about other users' prompts. Because prompt caching may cause privacy leakage, transparency around the caching policies of API providers is important. To this end, we develop and conduct statistical audits to detect prompt caching in real-world LLM API providers. We detect global cache sharing across users in seven API providers, including OpenAI, resulting in potential privacy leakage about users' prompts. Timing variations due to prompt caching can also result in leakage of information about model architecture. Namely, we find evidence that OpenAI's embedding model is a decoder-only Transformer, which was previously not publicly known.
comment: Accepted at ICML 2025
♻ ☆ Personalization of Large Language Models: A Survey
Personalization of Large Language Models (LLMs) has recently become increasingly important with a wide range of applications. Despite the importance and recent progress, most existing works on personalized LLMs have focused either entirely on (a) personalized text generation or (b) leveraging LLMs for personalization-related downstream applications, such as recommendation systems. In this work, we bridge the gap between these two separate main directions for the first time by introducing a taxonomy for personalized LLM usage and summarizing the key differences and challenges. We provide a formalization of the foundations of personalized LLMs that consolidates and expands notions of personalization of LLMs, defining and discussing novel facets of personalization, usage, and desiderata of personalized LLMs. We then unify the literature across these diverse fields and usage scenarios by proposing systematic taxonomies for the granularity of personalization, personalization techniques, datasets, evaluation methods, and applications of personalized LLMs. Finally, we highlight challenges and important open problems that remain to be addressed. By unifying and surveying recent research using the proposed taxonomies, we aim to provide a clear guide to the existing literature and different facets of personalization in LLMs, empowering both researchers and practitioners.
comment: Accepted at the Transactions on Machine Learning Research (TMLR) journal
♻ ☆ StreamUni: Achieving Streaming Speech Translation with a Unified Large Speech-Language Model
Streaming speech translation (StreamST) requires determining appropriate timing, known as policy, to generate translations while continuously receiving source speech inputs, balancing low latency with high translation quality. However, existing StreamST methods typically operate on sentence-level speech segments, referred to as simultaneous speech translation (SimulST). In practice, they require collaboration with segmentation models to accomplish StreamST, where the truncated speech segments constrain SimulST models to make policy decisions and generate translations based on limited contextual information. Moreover, SimulST models struggle to learn effective policies due to the complexity of speech inputs and cross-lingual generation. To address these challenges, we propose StreamUni, which achieves StreamST through a unified Large Speech-Language Model (LSLM). Specifically, StreamUni incorporates speech Chain-of-Thought (CoT) in guiding the LSLM to generate multi-stage outputs. Leveraging these multi-stage outputs, StreamUni simultaneously accomplishes speech segmentation, policy decision, and translation generation, completing StreamST without requiring massive policy-specific training. Additionally, we propose a streaming CoT training method that enhances low-latency policy decisions and generation capabilities using limited CoT data. Experiments demonstrate that our approach achieves state-of-the-art performance on StreamST tasks.
comment: The code is at https://github.com/ictnlp/StreamUni; The model is at https://huggingface.co/ICTNLP/StreamUni-Phi4
Information Retrieval 10
☆ Generative Cognitive Diagnosis
Cognitive diagnosis (CD) models latent cognitive states of human learners by analyzing their response patterns on diagnostic tests, serving as a crucial machine learning technique for educational assessment and evaluation. Traditional cognitive diagnosis models typically follow a transductive prediction paradigm that optimizes parameters to fit response scores and extract learner abilities. These approaches face significant limitations as they cannot perform instant diagnosis for new learners without computationally expensive retraining and produce diagnostic outputs with limited reliability. In this study, we introduces a novel generative diagnosis paradigm that fundamentally shifts CD from predictive to generative modeling, enabling inductive inference of cognitive states without parameter re-optimization. We propose two simple yet effective instantiations of this paradigm: Generative Item Response Theory (G-IRT) and Generative Neural Cognitive Diagnosis Model (G-NCDM), which achieve excellent performance improvements over traditional methods. The generative approach disentangles cognitive state inference from response prediction through a well-designed generation process that incorporates identifiability and monotonicity conditions. Extensive experiments on real-world datasets demonstrate the effectiveness of our methodology in addressing scalability and reliability challenges, especially $\times 100$ speedup for the diagnosis of new learners. Our framework opens new avenues for cognitive diagnosis applications in artificial intelligence, particularly for intelligent model evaluation and intelligent education systems. The code is available at https://github.com/CSLiJT/Generative-CD.git.
comment: Preprint; 15 pages, 12 figures
☆ Identifying Offline Metrics that Predict Online Impact: A Pragmatic Strategy for Real-World Recommender Systems RecSys 2025
A critical challenge in recommender systems is to establish reliable relationships between offline and online metrics that predict real-world performance. Motivated by recent advances in Pareto front approximation, we introduce a pragmatic strategy for identifying offline metrics that align with online impact. A key advantage of this approach is its ability to simultaneously serve multiple test groups, each with distinct offline performance metrics, in an online experiment controlled by a single model. The method is model-agnostic for systems with a neural network backbone, enabling broad applicability across architectures and domains. We validate the strategy through a large-scale online experiment in the field of session-based recommender systems on the OTTO e-commerce platform. The online experiment identifies significant alignments between offline metrics and real-word click-through rate, post-click conversion rate and units sold. Our strategy provides industry practitioners with a valuable tool for understanding offline-to-online metric relationships and making informed, data-driven decisions.
comment: This work was accepted for publication in the 19th ACM Conference on Recommender Systems (RecSys 2025). The final published version will be available at the ACM Digital Library
☆ Criteria-Based LLM Relevance Judgments ICTIR 2025
Relevance judgments are crucial for evaluating information retrieval systems, but traditional human-annotated labels are time-consuming and expensive. As a result, many researchers turn to automatic alternatives to accelerate method development. Among these, Large Language Models (LLMs) provide a scalable solution by generating relevance labels directly through prompting. However, prompting an LLM for a relevance label without constraints often results in not only incorrect predictions but also outputs that are difficult for humans to interpret. We propose the Multi-Criteria framework for LLM-based relevance judgments, decomposing the notion of relevance into multiple criteria--such as exactness, coverage, topicality, and contextual fit--to improve the robustness and interpretability of retrieval evaluations compared to direct grading methods. We validate this approach on three datasets: the TREC Deep Learning tracks from 2019 and 2020, as well as LLMJudge (based on TREC DL 2023). Our results demonstrate that Multi-Criteria judgments enhance the system ranking/leaderboard performance. Moreover, we highlight the strengths and limitations of this approach relative to direct grading approaches, offering insights that can guide the development of future automatic evaluation frameworks in information retrieval.
comment: 10 pages, 3 figures, accepted to ICTIR 2025
☆ Does UMBRELA Work on Other LLMs? SIGIR 2025
We reproduce the UMBRELA LLM Judge evaluation framework across a range of large language models (LLMs) to assess its generalizability beyond the original study. Our investigation evaluates how LLM choice affects relevance assessment accuracy, focusing on leaderboard rank correlation and per-label agreement metrics. Results demonstrate that UMBRELA with DeepSeek V3 obtains very comparable performance to GPT-4o (used in original work). For LLaMA-3.3-70B we obtain slightly lower performance, which further degrades with smaller LLMs.
comment: 9 pages, 2 figures, accepted to SIGIR 2025
☆ Dynamic Sparse Causal-Attention Temporal Networks for Interpretable Causality Discovery in Multivariate Time Series
Understanding causal relationships in multivariate time series (MTS) is essential for effective decision-making in fields such as finance and marketing, where complex dependencies and lagged effects challenge conventional analytical approaches. We introduce Dynamic Sparse Causal-Attention Temporal Networks for Interpretable Causality Discovery in MTS (DyCAST-Net), a novel architecture designed to enhance causal discovery by integrating dilated temporal convolutions and dynamic sparse attention mechanisms. DyCAST-Net effectively captures multiscale temporal dependencies through dilated convolutions while leveraging an adaptive thresholding strategy in its attention mechanism to eliminate spurious connections, ensuring both accuracy and interpretability. A statistical shuffle test validation further strengthens robustness by filtering false positives and improving causal inference reliability. Extensive evaluations on financial and marketing datasets demonstrate that DyCAST-Net consistently outperforms existing models such as TCDF, GCFormer, and CausalFormer. The model provides a more precise estimation of causal delays and significantly reduces false discoveries, particularly in noisy environments. Moreover, attention heatmaps offer interpretable insights, uncovering hidden causal patterns such as the mediated effects of advertising on consumer behavior and the influence of macroeconomic indicators on financial markets. Case studies illustrate DyCAST-Net's ability to detect latent mediators and lagged causal factors, making it particularly effective in high-dimensional, dynamic settings. The model's architecture enhanced by RMSNorm stabilization and causal masking ensures scalability and adaptability across diverse application domains
♻ ☆ Vision-Guided Chunking Is All You Need: Enhancing RAG with Multimodal Document Understanding
Retrieval-Augmented Generation (RAG) systems have revolutionized information retrieval and question answering, but traditional text-based chunking methods struggle with complex document structures, multi-page tables, embedded figures, and contextual dependencies across page boundaries. We present a novel multimodal document chunking approach that leverages Large Multimodal Models (LMMs) to process PDF documents in batches while maintaining semantic coherence and structural integrity. Our method processes documents in configurable page batches with cross-batch context preservation, enabling accurate handling of tables spanning multiple pages, embedded visual elements, and procedural content. We evaluate our approach on a curated dataset of PDF documents with manually crafted queries, demonstrating improvements in chunk quality and downstream RAG performance. Our vision-guided approach achieves better accuracy compared to traditional vanilla RAG systems, with qualitative analysis showing superior preservation of document structure and semantic coherence.
comment: 11 pages, 1 Figure, 1 Table
♻ ☆ Predictive Modeling: BIM Command Recommendation Based on Large-scale Usage Logs
The adoption of Building Information Modeling (BIM) and model-based design within the Architecture, Engineering, and Construction (AEC) industry has been hindered by the perception that using BIM authoring tools demands more effort than conventional 2D drafting. To enhance design efficiency, this paper proposes a BIM command recommendation framework that predicts the optimal next actions in real-time based on users' historical interactions. We propose a comprehensive filtering and enhancement method for large-scale raw BIM log data and introduce a novel command recommendation model. Our model builds upon the state-of-the-art Transformer backbones originally developed for large language models (LLMs), incorporating a custom feature fusion module, dedicated loss function, and targeted learning strategy. In a case study, the proposed method is applied to over 32 billion rows of real-world log data collected globally from the BIM authoring software Vectorworks. Experimental results demonstrate that our method can learn universal and generalizable modeling patterns from anonymous user interaction sequences across different countries, disciplines, and projects. When generating recommendations for the next command, our approach achieves a Recall@10 of approximately 84%. The code is available at: https://github.com/dcy0577/BIM-Command-Recommendation.git
comment: Advanced Engineering Informatics
♻ ☆ Exploring the Effect of Context-Awareness and Popularity Calibration on Popularity Bias in POI Recommendations RecSys 2025
Point-of-interest (POI) recommender systems help users discover relevant locations, but their effectiveness is often compromised by popularity bias, which disadvantages less popular, yet potentially meaningful places. This paper addresses this challenge by evaluating the effectiveness of context-aware models and calibrated popularity techniques as strategies for mitigating popularity bias. Using four real-world POI datasets (Brightkite, Foursquare, Gowalla, and Yelp), we analyze the individual and combined effects of these approaches on recommendation accuracy and popularity bias. Our results reveal that context-aware models cannot be considered a uniform solution, as the models studied exhibit divergent impacts on accuracy and bias. In contrast, calibration techniques can effectively align recommendation popularity with user preferences, provided there is a careful balance between accuracy and bias mitigation. Notably, the combination of calibration and context-awareness yields recommendations that balance accuracy and close alignment with the users' popularity profiles, i.e., popularity calibration.
comment: Accepted at RecSys 2025, DOI: https://doi.org/10.1145/3705328.3748017
♻ ☆ MoRE: A Mixture of Reflectors Framework for Large Language Model-Based Sequential Recommendation RecSys'25
Large language models (LLMs) have emerged as a cutting-edge approach in sequential recommendation, leveraging historical interactions to model dynamic user preferences. Current methods mainly focus on learning processed recommendation data in the form of sequence-to-sequence text. While effective, they exhibit three key limitations: 1) failing to decouple intra-user explicit features (e.g., product titles) from implicit behavioral patterns (e.g., brand loyalty) within interaction histories; 2) underutilizing cross-user collaborative filtering (CF) signals; and 3) relying on inefficient reflection update strategies. To address this, We propose MoRE (Mixture of REflectors), which introduces three perspective-aware offline reflection processes to address these gaps. This decomposition directly resolves Challenges 1 (explicit/implicit ambiguity) and 2 (CF underutilization). Furthermore, MoRE's meta-reflector employs a self-improving strategy and a dynamic selection mechanism (Challenge 3) to adapt to evolving user preferences. First, two intra-user reflectors decouple explicit and implicit patterns from a user's interaction sequence, mimicking traditional recommender systems' ability to distinguish surface-level and latent preferences. A third cross-user reflector captures CF signals by analyzing user similarity patterns from multiple users' interactions. To optimize reflection quality, MoRE's meta-reflector employs a offline self-improving strategy that evaluates reflection impacts through comparisons of presence/absence and iterative refinement of old/new versions, with a online contextual bandit mechanism dynamically selecting the optimal perspective for recommendation for each user. Code: https://github.com/E-qin/MoRE-Rec.
comment: First 2 authors contributes equally to this work, accepted by RecSys'25 spotlight oral. Corresponding author is Weijie Yu(yu@uibe.edu.cn)
♻ ☆ LLM-Driven Usefulness Labeling for IR Evaluation
In the information retrieval (IR) domain, evaluation plays a crucial role in optimizing search experiences and supporting diverse user intents. In the recent LLM era, research has been conducted to automate document relevance labels, as these labels have traditionally been assigned by crowd-sourced workers - a process that is both time and consuming and costly. This study focuses on LLM-generated usefulness labels, a crucial evaluation metric that considers the user's search intents and task objectives, an aspect where relevance falls short. Our experiment utilizes task-level, query-level, and document-level features along with user search behavior signals, which are essential in defining the usefulness of a document. Our research finds that (i) pre-trained LLMs can generate moderate usefulness labels by understanding the comprehensive search task session, (ii) pre-trained LLMs perform better judgement in short search sessions when provided with search session contexts. Additionally, we investigated whether LLMs can capture the unique divergence between relevance and usefulness, along with conducting an ablation study to identify the most critical metrics for accurate usefulness label generation. In conclusion, this work explores LLM-generated usefulness labels by evaluating critical metrics and optimizing for practicality in real-world settings.
Multimedia 4
☆ KEN: Knowledge Augmentation and Emotion Guidance Network for Multimodal Fake News Detection ACM MM 2025
In recent years, the rampant spread of misinformation on social media has made accurate detection of multimodal fake news a critical research focus. However, previous research has not adequately understood the semantics of images, and models struggle to discern news authenticity with limited textual information. Meanwhile, treating all emotional types of news uniformly without tailored approaches further leads to performance degradation. Therefore, we propose a novel Knowledge Augmentation and Emotion Guidance Network (KEN). On the one hand, we effectively leverage LVLM's powerful semantic understanding and extensive world knowledge. For images, the generated captions provide a comprehensive understanding of image content and scenes, while for text, the retrieved evidence helps break the information silos caused by the closed and limited text and context. On the other hand, we consider inter-class differences between different emotional types of news through balanced learning, achieving fine-grained modeling of the relationship between emotional types and authenticity. Extensive experiments on two real-world datasets demonstrate the superiority of our KEN.
comment: Accepted by ACM MM 2025
♻ ☆ High-Quality Live Video Streaming via Transcoding Time Prediction and Preset Selection
Video streaming often requires transcoding content into different resolutions and bitrates to match the recipient's internet speed and screen capabilities. Video encoders like x264 offer various presets, each with different tradeoffs between transcoding time and rate-distortion performance. Choosing the best preset for video transcoding is difficult, especially for live streaming, as trying all the presets and choosing the best one is not feasible. One solution is to predict each preset's transcoding time and select the preset that ensures the highest quality while adhering to live streaming time constraints. Prediction of video transcoding time is also critical in minimizing streaming delays, deploying resource management algorithms, and load balancing. We propose a learning-based framework for predicting the transcoding time of videos across various presets. Our predictor's features for video transcoding time prediction are derived directly from the ingested stream, primarily from the header or metadata. As a result, only minimal additional delay is incurred for feature extraction, rendering our approach ideal for live-streaming applications. We evaluated our learning-based transcoding time prediction using a dataset of videos. The results demonstrate that our framework can accurately predict the transcoding time for different presets, with a mean absolute percentage error (MAPE) of nearly 5.0%. Leveraging these predictions, we then select the most suitable transcoding preset for live video streaming. Utilizing our transcoding time prediction-based preset selection improved Peak Signal-to-Noise Ratio (PSNR) of up to 5 dB.
comment: After further review, we found major flaws in the paper that need extensive revision
♻ ☆ Hear-Your-Click: Interactive Object-Specific Video-to-Audio Generation
Video-to-audio (V2A) generation shows great potential in fields such as film production. Despite significant advances, current V2A methods relying on global video information struggle with complex scenes and generating audio tailored to specific objects. To address these limitations, we introduce Hear-Your-Click, an interactive V2A framework enabling users to generate sounds for specific objects by clicking on the frame. To achieve this, we propose Object-aware Contrastive Audio-Visual Fine-tuning (OCAV) with a Mask-guided Visual Encoder (MVE) to obtain object-level visual features aligned with audio. Furthermore, we tailor two data augmentation strategies, Random Video Stitching (RVS) and Mask-guided Loudness Modulation (MLM), to enhance the model's sensitivity to segmented objects. To measure audio-visual correspondence, we designed a new evaluation metric, the CAV score. Extensive experiments demonstrate that our framework offers more precise control and improves generation performance across various metrics. Project Page: https://github.com/SynapGrid/Hear-Your-Click
♻ ☆ MEDTalk: Multimodal Controlled 3D Facial Animation with Dynamic Emotions by Disentangled Embedding
Audio-driven emotional 3D facial animation aims to generate synchronized lip movements and vivid facial expressions. However, most existing approaches focus on static and predefined emotion labels, limiting their diversity and naturalness. To address these challenges, we propose MEDTalk, a novel framework for fine-grained and dynamic emotional talking head generation. Our approach first disentangles content and emotion embedding spaces from motion sequences using a carefully designed cross-reconstruction process, enabling independent control over lip movements and facial expressions. Beyond conventional audio-driven lip synchronization, we integrate audio and speech text, predicting frame-wise intensity variations and dynamically adjusting static emotion features to generate realistic emotional expressions. Furthermore, to enhance control and personalization, we incorporate multimodal inputs-including text descriptions and reference expression images-to guide the generation of user-specified facial expressions. With MetaHuman as the priority, our generated results can be conveniently integrated into the industrial production pipeline.
comment: 11 pages, 8 figures
Computation and Language 26
☆ DATE-LM: Benchmarking Data Attribution Evaluation for Large Language Models
Data attribution methods quantify the influence of training data on model outputs and are becoming increasingly relevant for a wide range of LLM research and applications, including dataset curation, model interpretability, data valuation. However, there remain critical gaps in systematic LLM-centric evaluation of data attribution methods. To this end, we introduce DATE-LM (Data Attribution Evaluation in Language Models), a unified benchmark for evaluating data attribution methods through real-world LLM applications. DATE-LM measures attribution quality through three key tasks -- training data selection, toxicity/bias filtering, and factual attribution. Our benchmark is designed for ease of use, enabling researchers to configure and run large-scale evaluations across diverse tasks and LLM architectures. Furthermore, we use DATE-LM to conduct a large-scale evaluation of existing data attribution methods. Our findings show that no single method dominates across all tasks, data attribution methods have trade-offs with simpler baselines, and method performance is sensitive to task-specific evaluation design. Finally, we release a public leaderboard for quick comparison of methods and to facilitate community engagement. We hope DATE-LM serves as a foundation for future data attribution research in LLMs.
☆ ZipVoice-Dialog: Non-Autoregressive Spoken Dialogue Generation with Flow Matching
Generating spoken dialogue is more challenging than monologue text-to-speech (TTS) due to the need for realistic turn-taking and distinct speaker timbres. Existing spoken dialogue generation models, being auto-regressive, suffer from slow and unstable inference. To overcome these limitations, we introduce ZipVoice-Dialog, a non-autoregressive zero-shot spoken dialogue generation model built upon flow matching. Key designs include: 1) speaker-turn embeddings for precise speaker turn-taking; 2) a curriculum learning strategy for stable speech-text alignment; 3) specialized strategies to enable stereo dialogue generation. Additionally, recognizing the lack of open-source large-scale spoken dialogue datasets, we curated OpenDialog, a 6.8k-hour spoken dialogue dataset from in-the-wild speech data. Furthermore, we established a benchmark to comprehensively evaluate various models. Experimental results demonstrate that ZipVoice-Dialog achieves superior performance in intelligibility, speaker turn-taking accuracy, speaker similarity, and inference speed. Our codes, model checkpoints, demo samples, and the OpenDialog dataset are all publicly available at https://github.com/k2-fsa/ZipVoice.
☆ Voice Conversion for Lombard Speaking Style with Implicit and Explicit Acoustic Feature Conditioning
Text-to-Speech (TTS) systems in Lombard speaking style can improve the overall intelligibility of speech, useful for hearing loss and noisy conditions. However, training those models requires a large amount of data and the Lombard effect is challenging to record due to speaker and noise variability and tiring recording conditions. Voice conversion (VC) has been shown to be a useful augmentation technique to train TTS systems in the absence of recorded data from the target speaker in the target speaking style. In this paper, we are concerned with Lombard speaking style transfer. Our goal is to convert speaker identity while preserving the acoustic attributes that define the Lombard speaking style. We compare voice conversion models with implicit and explicit acoustic feature conditioning. We observe that our proposed implicit conditioning strategy achieves an intelligibility gain comparable to the model conditioned on explicit acoustic features, while also preserving speaker similarity.
comment: Presented at Clarity Challenge 2023
☆ ClaritySpeech: Dementia Obfuscation in Speech
Dementia, a neurodegenerative disease, alters speech patterns, creating communication barriers and raising privacy concerns. Current speech technologies, such as automatic speech transcription (ASR), struggle with dementia and atypical speech, further challenging accessibility. This paper presents a novel dementia obfuscation in speech framework, ClaritySpeech, integrating ASR, text obfuscation, and zero-shot text-to-speech (TTS) to correct dementia-affected speech while preserving speaker identity in low-data environments without fine-tuning. Results show a 16% and 10% drop in mean F1 score across various adversarial settings and modalities (audio, text, fusion) for ADReSS and ADReSSo, respectively, maintaining 50% speaker similarity. We also find that our system improves WER (from 0.73 to 0.08 for ADReSS and 0.15 for ADReSSo) and speech quality from 1.65 to ~2.15, enhancing privacy and accessibility.
comment: Accepted at Interspeech 2025
Prompt4Trust: A Reinforcement Learning Prompt Augmentation Framework for Clinically-Aligned Confidence Calibration in Multimodal Large Language Models ICCV 2025
Multimodal large language models (MLLMs) hold considerable promise for applications in healthcare. However, their deployment in safety-critical settings is hindered by two key limitations: (i) sensitivity to prompt design, and (ii) a tendency to generate incorrect responses with high confidence. As clinicians may rely on a model's stated confidence to gauge the reliability of its predictions, it is especially important that when a model expresses high confidence, it is also highly accurate. We introduce Prompt4Trust, the first reinforcement learning (RL) framework for prompt augmentation targeting confidence calibration in MLLMs. A lightweight LLM is trained to produce context-aware auxiliary prompts that guide a downstream task MLLM to generate responses in which the expressed confidence more accurately reflects predictive accuracy. Unlike conventional calibration techniques, Prompt4Trust specifically prioritizes aspects of calibration most critical for safe and trustworthy clinical decision-making. Beyond improvements driven by this clinically motivated calibration objective, our proposed method also improves task accuracy, achieving state-of-the-art medical visual question answering (VQA) performance on the PMC-VQA benchmark, which is composed of multiple-choice questions spanning diverse medical imaging modalities. Moreover, our framework trained with a small downstream task MLLM showed promising zero-shot generalization to larger MLLMs in our experiments, suggesting the potential for scalable calibration without the associated computational costs. This work demonstrates the potential of automated yet human-aligned prompt engineering for improving the the trustworthiness of MLLMs in safety critical settings. Our codebase can be found at https://github.com/xingbpshen/vccrl-llm.
comment: Preprint version. The peer-reviewed version of this paper has been accepted to ICCV 2025 Workshop CVAMD
☆ Psychology-Driven Enhancement of Humour Translation
Humour translation plays a vital role as a bridge between different cultures, fostering understanding and communication. Although most existing Large Language Models (LLMs) are capable of general translation tasks, these models still struggle with humour translation, which is especially reflected through linguistic interference and lacking humour in translated text. In this paper, we propose a psychology-inspired Humour Decomposition Mechanism (HDM) that utilises Chain-of-Thought (CoT) to imitate the ability of the human thought process, stimulating LLMs to optimise the readability of translated humorous texts. Moreover, we integrate humour theory in HDM to further enhance the humorous elements in the translated text. Our automatic evaluation experiments on open-source humour datasets demonstrate that our method significantly improves the quality of humour translation, yielding average gains of 7.75\% in humour, 2.81\% in fluency, and 6.13\% in coherence of the generated text.
☆ Swa-bhasha Resource Hub: Romanized Sinhala to Sinhala Transliteration Systems and Data Resources
The Swa-bhasha Resource Hub provides a comprehensive collection of data resources and algorithms developed for Romanized Sinhala to Sinhala transliteration between 2020 and 2025. These resources have played a significant role in advancing research in Sinhala Natural Language Processing (NLP), particularly in training transliteration models and developing applications involving Romanized Sinhala. The current openly accessible data sets and corresponding tools are made publicly available through this hub. This paper presents a detailed overview of the resources contributed by the authors and includes a comparative analysis of existing transliteration applications in the domain.
comment: 13 pages, 3 Tables, 3 figures
☆ MetaClimage: A novel database of visual metaphors related to Climate Change, with costs and benefits analysis
Visual metaphors of climate change (e.g., melting glaciers depicted as a melting ice grenade) are regarded as valuable tools for addressing the complexity of environmental challenges. However, few studies have examined their impact on communication, also due to scattered availability of material. Here, we present a novel database of Metaphors of Climate Change in Images (MetaClimage) https://doi.org/10.5281/zenodo.15861012, paired with literal images and enriched with human ratings. For each image, we collected values of difficulty, efficacy, artistic quality, and emotional arousal from human rating, as well as number of tags generated by participants to summarize the message. Semantic and emotion variables were further derived from the tags via Natural Language Processing. Visual metaphors were rated as more difficult to understand, yet more aesthetically pleasant than literal images, but did not differ in efficacy and arousal. The latter for visual metaphors, however, was higher in participants with higher Need For Cognition. Furthermore, visual metaphors received more tags, often referring to entities not depicted in the image, and elicited words with more positive valence and greater dominance than literal images. These results evidence the greater cognitive load of visual metaphors, which nevertheless might induce positive effects such as deeper cognitive elaboration and abstraction compared to literal stimuli. Furthermore, while they are not deemed as more effective and arousing, visual metaphors seem to generate superior aesthetic appreciation and a more positively valenced experience. Overall, this study contributes to understanding the impact of visual metaphors of climate change both by offering a database for future research and by elucidating a cost-benefit trade-off to take into account when shaping environmental communication.
comment: 27 pages, 5 figures
☆ Banzhida: Advancing Large Language Models for Tibetan with Curated Data and Continual Pre-Training
Large language models have achieved remarkable progress across many languages. However, Tibetan, as a representative low-resource language, is particularly underrepresented in existing models due to the scarcity of high-quality training corpora. To address this gap, we curate the largest Tibetan pre-training corpus to date, aggregating data from diverse sources and applying a dedicated data cleaning and processing pipeline tailored for Tibetan. With the curated data, we continue pre/post-training a multilingual base model into Banzhida, a multilingual large language model that advances generative AI for Tibetan. To evaluate the Tibetan capabilities of the model, we create new high-quality Tibetan benchmarks, and complement them with existing public benchmarks. Experimental results demonstrate that Banzhida consistently and significantly outperforms both open-source models of similar scale and Tibetan-tailored models across a wide range of tasks.
♻ ☆ Large Language Models as Neurolinguistic Subjects: Discrepancy between Performance and Competence
This study investigates the linguistic understanding of Large Language Models (LLMs) regarding signifier (form) and signified (meaning) by distinguishing two LLM assessment paradigms: psycholinguistic and neurolinguistic. Traditional psycholinguistic evaluations often reflect statistical rules that may not accurately represent LLMs' true linguistic competence. We introduce a neurolinguistic approach, utilizing a novel method that combines minimal pair and diagnostic probing to analyze activation patterns across model layers. This method allows for a detailed examination of how LLMs represent form and meaning, and whether these representations are consistent across languages. We found: (1) Psycholinguistic and neurolinguistic methods reveal that language performance and competence are distinct; (2) Direct probability measurement may not accurately assess linguistic competence; (3) Instruction tuning won't change much competence but improve performance; (4) LLMs exhibit higher competence and performance in form compared to meaning. Additionally, we introduce new conceptual minimal pair datasets for Chinese (COMPS-ZH) and German (COMPS-DE), complementing existing English datasets.
♻ ☆ A Survey of Automatic Prompt Optimization with Instruction-focused Heuristic-based Search Algorithm ACL 2025
Recent advances in Large Language Models have led to remarkable achievements across a variety of Natural Language Processing tasks, making prompt engineering increasingly central to guiding model outputs. While manual methods can be effective, they typically rely on intuition and do not automatically refine prompts over time. In contrast, automatic prompt optimization employing heuristic-based search algorithms can systematically explore and improve prompts with minimal human oversight. This survey proposes a comprehensive taxonomy of these methods, categorizing them by where optimization occurs, what is optimized, what criteria drive the optimization, which operators generate new prompts, and which iterative search algorithms are applied. We further highlight specialized datasets and tools that support and accelerate automated prompt refinement. We conclude by discussing key open challenges pointing toward future opportunities for more robust and versatile LLM applications.
comment: Accepted to ACL 2025
♻ ☆ Single Word Change is All You Need: Using LLMs to Create Synthetic Training Examples for Text Classifiers
In text classification, creating an adversarial example means subtly perturbing a few words in a sentence without changing its meaning, causing it to be misclassified by a classifier. A concerning observation is that a significant portion of adversarial examples generated by existing methods change only one word. This single-word perturbation vulnerability represents a significant weakness in classifiers, which malicious users can exploit to efficiently create a multitude of adversarial examples. This paper studies this problem and makes the following key contributions: (1) We introduce a novel metric $\rho$ to quantitatively assess a classifier's robustness against single-word perturbation. (2) We present the SP-Attack, designed to exploit the single-word perturbation vulnerability, achieving a higher attack success rate, better preserving sentence meaning, while reducing computation costs compared to state-of-the-art adversarial methods. (3) We propose SP-Defense, which aims to improve \r{ho} by applying data augmentation in learning. Experimental results on 4 datasets and BERT and distilBERT classifiers show that SP-Defense improves $\rho$ by 14.6% and 13.9% and decreases the attack success rate of SP-Attack by 30.4% and 21.2% on two classifiers respectively, and decreases the attack success rate of existing attack methods that involve multiple-word perturbations.
♻ ☆ SEE: Strategic Exploration and Exploitation for Cohesive In-Context Prompt Optimization ACL 2025
Designing optimal prompts for Large Language Models (LLMs) is a complicated and resource-intensive task, often requiring substantial human expertise and effort. Existing approaches typically separate the optimization of prompt instructions and in-context learning examples, leading to incohesive prompts that are defined and represented by suboptimal task performance. To overcome these challenges, we propose a novel Cohesive In-Context Prompt Optimization framework that refines both prompt instructions and examples. However, formulating such an optimization in the discrete and high-dimensional space of natural language poses significant challenges in both convergence and computational efficiency. To address these issues, we introduce SEE, a scalable and efficient prompt optimization framework that adopts metaheuristic optimization principles and strategically balances exploration and exploitation to enhance optimization performance and achieve efficient convergence. SEE features a quad-phased design that alternates between global traversal (exploration) and local optimization (exploitation) and adaptively chooses LLM operators during the optimization process. We have conducted a comprehensive evaluation across 35 benchmark tasks, and SEE significantly outperforms state-of-the-art baseline methods by a large margin, achieving an average performance gain of 13.94 while reducing computational costs by 58.67.
comment: Accepted to ACL 2025 (Main Conference)
♻ ☆ Supposedly Equivalent Facts That Aren't? Entity Frequency in Pre-training Induces Asymmetry in LLMs
Understanding and mitigating hallucinations in Large Language Models (LLMs) is crucial for ensuring reliable content generation. While previous research has primarily focused on "when" LLMs hallucinate, our work explains "why" and directly links model behaviour to the pre-training data that forms their prior knowledge. Specifically, we demonstrate that an asymmetry exists in the recognition of logically equivalent facts, which can be attributed to frequency discrepancies of entities appearing as subjects versus objects. Given that most pre-training datasets are inaccessible, we leverage the fully open-source OLMo series by indexing its Dolma dataset to estimate entity frequencies. Using relational facts (represented as triples) from Wikidata5M, we construct probing datasets to isolate this effect. Our experiments reveal that facts with a high-frequency subject and a low-frequency object are better recognised than their inverse, despite their logical equivalence. The pattern reverses in low-to-high frequency settings, and no statistically significant asymmetry emerges when both entities are high-frequency. These findings highlight the influential role of pre-training data in shaping model predictions and provide insights for inferring the characteristics of pre-training data in closed or partially closed LLMs.
comment: Accepted at COLM 2025
♻ ☆ MedGemma Technical Report
Artificial intelligence (AI) has significant potential in healthcare applications, but its training and deployment faces challenges due to healthcare's diverse data, complex tasks, and the need to preserve privacy. Foundation models that perform well on medical tasks and require less task-specific tuning data are critical to accelerate the development of healthcare AI applications. We introduce MedGemma, a collection of medical vision-language foundation models based on Gemma 3 4B and 27B. MedGemma demonstrates advanced medical understanding and reasoning on images and text, significantly exceeding the performance of similar-sized generative models and approaching the performance of task-specific models, while maintaining the general capabilities of the Gemma 3 base models. For out-of-distribution tasks, MedGemma achieves 2.6-10% improvement on medical multimodal question answering, 15.5-18.1% improvement on chest X-ray finding classification, and 10.8% improvement on agentic evaluations compared to the base models. Fine-tuning MedGemma further improves performance in subdomains, reducing errors in electronic health record information retrieval by 50% and reaching comparable performance to existing specialized state-of-the-art methods for pneumothorax classification and histopathology patch classification. We additionally introduce MedSigLIP, a medically-tuned vision encoder derived from SigLIP. MedSigLIP powers the visual understanding capabilities of MedGemma and as an encoder achieves comparable or better performance than specialized medical image encoders. Taken together, the MedGemma collection provides a strong foundation of medical image and text capabilities, with potential to significantly accelerate medical research and development of downstream applications. The MedGemma collection, including tutorials and model weights, can be found at https://goo.gle/medgemma.
♻ ☆ BEExformer: A Fast Inferencing Binarized Transformer with Early Exits
Large Language Models (LLMs) based on transformers achieve cutting-edge results on a variety of applications. However, their enormous size and processing requirements hinder deployment on constrained resources. To enhance efficiency, binarization and Early Exit (EE) have proved to be effective solutions. However, binarization may lead to performance loss as reduced precision affects gradient estimation and parameter updates. Besides, research on EE mechanisms is still in its early stages. To address these challenges, we introduce Binarized Early Exit Transformer (BEExformer), the first-ever selective learning-based transformer integrating Binarization-Aware Training (BAT) with EE for efficient and fast textual inference. Each transformer block has an integrated Selective-Learn Forget Network (SLFN) to enhance contextual retention while eliminating irrelevant information. The BAT employs a differentiable second-order approximation to the sign function, enabling gradient computation that captures both the sign and magnitude of the weights. This aids in 21.30 times reduction in model size. The EE mechanism hinges on fractional reduction in entropy among intermediate transformer blocks with soft-routing loss estimation. This accelerates inference by reducing FLOPs by 52.08% and even improves accuracy by 2.89% by resolving the "overthinking" problem inherent in deep networks. Extensive evaluation through comparison with the SOTA methods and various ablations across six datasets covering multiple NLP tasks demonstrates its Pareto-optimal performance-efficiency trade-off.
comment: This revised manuscript includes 18 pages, 17 figures, and 6 tables. Methodology and results sections have been improved for clarity and depth, incorporating additional comparisons, ablations, and a new evaluation dataset. A few relevant references were added, and overall organization refined for better readability
♻ ☆ Perspective Dial: Measuring Perspective of Text and Guiding LLM Outputs
Large language models (LLMs) are used in a variety of mission-critical roles. Due to the rapidly developing nature of LLMs, there is a lack of quantifiable understanding of the bias and perspective associated with LLM output. Inspired by this need, this paper considers the broader issue of perspective or viewpoint of general text and perspective control of large-language model (LLM) output. Perspective-Dial consists of two main components: a (1) metric space, dubbed Perspective Space, that enables quantitative measurements of different perspectives regarding a topic, and the use of (2) Systematic Prompt Engineering that utilizes greedy-coordinate descent to control LLM output perspective based on measurement feedback from the Perspective Space. The empirical nature of the approach allows progress to side step a principled understanding of perspective or bias -- effectively quantifying and adjusting outputs for a variety of topics. Potential applications include detection, tracking and mitigation of LLM bias, narrative detection, sense making and tracking in public discourse, and debate bot advocating given perspective.
comment: 7 pages, 5 main pages of text, 5 figures, 2 tables. Research work performed at CACI INTL INC
♻ ☆ Watermarking Degrades Alignment in Language Models: Analysis and Mitigation ICLR 2025
Watermarking techniques for large language models (LLMs) can significantly impact output quality, yet their effects on truthfulness, safety, and helpfulness remain critically underexamined. This paper presents a systematic analysis of how two popular watermarking approaches-Gumbel and KGW-affect these core alignment properties across four aligned LLMs. Our experiments reveal two distinct degradation patterns: guard attenuation, where enhanced helpfulness undermines model safety, and guard amplification, where excessive caution reduces model helpfulness. These patterns emerge from watermark-induced shifts in token distribution, surfacing the fundamental tension that exists between alignment objectives. To mitigate these degradations, we propose Alignment Resampling (AR), an inference-time sampling method that uses an external reward model to restore alignment. We establish a theoretical lower bound on the improvement in expected reward score as the sample size is increased and empirically demonstrate that sampling just 2-4 watermarked generations effectively recovers or surpasses baseline (unwatermarked) alignment scores. To overcome the limited response diversity of standard Gumbel watermarking, our modified implementation sacrifices strict distortion-freeness while maintaining robust detectability, ensuring compatibility with AR. Experimental results confirm that AR successfully recovers baseline alignment in both watermarking approaches, while maintaining strong watermark detectability. This work reveals the critical balance between watermark strength and model alignment, providing a simple inference-time solution to responsibly deploy watermarked LLMs in practice.
comment: Published at the 1st Workshop on GenAI Watermarking (ICLR 2025). Code: https://github.com/dapurv5/alignmark
♻ ☆ LLM Agents Are the Antidote to Walled Gardens
While the Internet's core infrastructure was designed to be open and universal, today's application layer is dominated by closed, proprietary platforms. Open and interoperable APIs require significant investment, and market leaders have little incentive to enable data exchange that could erode their user lock-in. We argue that LLM-based agents fundamentally disrupt this status quo. Agents can automatically translate between data formats and interact with interfaces designed for humans: this makes interoperability dramatically cheaper and effectively unavoidable. We name this shift universal interoperability: the ability for any two digital services to exchange data seamlessly using AI-mediated adapters. Universal interoperability undermines monopolistic behaviours and promotes data portability. However, it can also lead to new security risks and technical debt. Our position is that the ML community should embrace this development while building the appropriate frameworks to mitigate the downsides. By acting now, we can harness AI to restore user freedom and competitive markets without sacrificing security.
♻ ☆ Evaluation of Attribution Bias in Generator-Aware Retrieval-Augmented Large Language Models ACL 2025
Attributing answers to source documents is an approach used to enhance the verifiability of a model's output in retrieval augmented generation (RAG). Prior work has mainly focused on improving and evaluating the attribution quality of large language models (LLMs) in RAG, but this may come at the expense of inducing biases in the attribution of answers. We define and examine two aspects in the evaluation of LLMs in RAG pipelines, namely attribution sensitivity and bias with respect to authorship information. We explicitly inform an LLM about the authors of source documents, instruct it to attribute its answers, and analyze (i) how sensitive the LLM's output is to the author of source documents, and (ii) whether the LLM exhibits a bias towards human-written or AI-generated source documents. We design an experimental setup in which we use counterfactual evaluation to study three LLMs in terms of their attribution sensitivity and bias in RAG pipelines. Our results show that adding authorship information to source documents can significantly change the attribution quality of LLMs by 3% to 18%. Moreover, we show that LLMs can have an attribution bias towards explicit human authorship, which can serve as a competing hypothesis for findings of prior work that shows that LLM-generated content may be preferred over human-written contents. Our findings indicate that metadata of source documents can influence LLMs' trust, and how they attribute their answers. Furthermore, our research highlights attribution bias and sensitivity as a novel aspect of brittleness in LLMs.
comment: Accepted at ACL 2025 (Findings)
♻ ☆ Disambiguate First, Parse Later: Generating Interpretations for Ambiguity Resolution in Semantic Parsing ACL 2025
Handling ambiguity and underspecification is an important challenge in natural language interfaces, particularly for tasks like text-to-SQL semantic parsing. We propose a modular approach that resolves ambiguity using natural language interpretations before mapping these to logical forms (e.g., SQL queries). Although LLMs excel at parsing unambiguous utterances, they show strong biases for ambiguous ones, typically predicting only preferred interpretations. We constructively exploit this bias to generate an initial set of preferred disambiguations and then apply a specialized infilling model to identify and generate missing interpretations. To train the infilling model, we introduce an annotation method that uses SQL execution to validate different meanings. Our approach improves interpretation coverage and generalizes across datasets with different annotation styles, database structures, and ambiguity types.
comment: Findings of ACL 2025
♻ ☆ Planted in Pretraining, Swayed by Finetuning: A Case Study on the Origins of Cognitive Biases in LLMs
Large language models (LLMs) exhibit cognitive biases -- systematic tendencies of irrational decision-making, similar to those seen in humans. Prior work has found that these biases vary across models and can be amplified by instruction tuning. However, it remains unclear if these differences in biases stem from pretraining, finetuning, or even random noise due to training stochasticity. We propose a two-step causal experimental approach to disentangle these factors. First, we finetune models multiple times using different random seeds to study how training randomness affects over $30$ cognitive biases. Second, we introduce \emph{cross-tuning} -- swapping instruction datasets between models to isolate bias sources. This swap uses datasets that led to different bias patterns, directly testing whether biases are dataset-dependent. Our findings reveal that while training randomness introduces some variability, biases are mainly shaped by pretraining: models with the same pretrained backbone exhibit more similar bias patterns than those sharing only finetuning data. These insights suggest that understanding biases in finetuned models requires considering their pretraining origins beyond finetuning effects. This perspective can guide future efforts to develop principled strategies for evaluating and mitigating bias in LLMs.
comment: CoLM 2025
♻ ☆ Towards Pareto Optimal Throughput in Small Language Model Serving
Large language models (LLMs) have revolutionized the state-of-the-art of many different natural language processing tasks. Although serving LLMs is computationally and memory demanding, the rise of Small Language Models (SLMs) offers new opportunities for resource-constrained users, who now are able to serve small models with cutting-edge performance. In this paper, we present a set of experiments designed to benchmark SLM inference at performance and energy levels. Our analysis provides a new perspective in serving, highlighting that the small memory footprint of SLMs allows for reaching the Pareto-optimal throughput within the resource capacity of a single accelerator. In this regard, we present an initial set of findings demonstrating how model replication can effectively improve resource utilization for serving SLMs.
comment: Revised version of the paper published at EuroMLSys'24
♻ ☆ Feature Extraction and Steering for Enhanced Chain-of-Thought Reasoning in Language Models
Large Language Models (LLMs) demonstrate the ability to solve reasoning and mathematical problems using the Chain-of-Thought (CoT) technique. Expanding CoT length, as seen in models such as DeepSeek-R1, significantly enhances this reasoning for complex problems, but requires costly and high-quality long CoT data and fine-tuning. This work, inspired by the deep thinking paradigm of DeepSeek-R1, utilizes a steering technique to enhance the reasoning ability of an LLM without external datasets. Our method first employs Sparse Autoencoders (SAEs) to extract interpretable features from vanilla CoT. These features are then used to steer the LLM's internal states during generation. Recognizing that many LLMs do not have corresponding pre-trained SAEs, we further introduce a novel SAE-free steering algorithm, which directly computes steering directions from the residual activations of an LLM, obviating the need for an explicit SAE. Experimental results demonstrate that both our SAE-based and subsequent SAE-free steering algorithms significantly enhance the reasoning capabilities of LLMs.
♻ ☆ Exploring Gender Bias Beyond Occupational Titles
In this work, we investigate the correlation between gender and contextual biases, focusing on elements such as action verbs, object nouns, and particularly on occupations. We introduce a novel dataset, GenderLexicon, and a framework that can estimate contextual bias and its related gender bias. Our model can interpret the bias with a score and thus improve the explainability of gender bias. Also, our findings confirm the existence of gender biases beyond occupational stereotypes. To validate our approach and demonstrate its effectiveness, we conduct evaluations on five diverse datasets, including a Japanese dataset.
comment: Work in progress
♻ ☆ An In-depth Evaluation of Large Language Models in Sentence Simplification with Error-based Human Assessment
Recent studies have used both automatic metrics and human evaluations to assess the simplification abilities of LLMs. However, the suitability of existing evaluation methodologies for LLMs remains in question. First, the suitability of current automatic metrics on LLMs' simplification evaluation is still uncertain. Second, current human evaluation approaches in sentence simplification often fall into two extremes: they are either too superficial, failing to offer a clear understanding of the models' performance, or overly detailed, making the annotation process complex and prone to inconsistency, which in turn affects the evaluation's reliability. To address these problems, this study provides in-depth insights into LLMs' performance while ensuring the reliability of the evaluation. We design an error-based human annotation framework to assess the LLMs' simplification capabilities. We select both closed-source and open-source LLMs, including GPT-4, Qwen2.5-72B, and Llama-3.2-3B. We believe that these models offer a representative selection across large, medium, and small sizes of LLMs. Results show that LLMs generally generate fewer erroneous simplification outputs compared to the previous state-of-the-art. However, LLMs have their limitations, as seen in GPT-4's and Qwen2.5-72B's struggle with lexical paraphrasing. Furthermore, we conduct meta-evaluations on widely used automatic metrics using our human annotations. We find that these metrics lack sufficient sensitivity to assess the overall high-quality simplifications, particularly those generated by high-performance LLMs.
comment: Accepted by ACM Transactions on Intelligent Systems and Technology. Our human evaluation corpus is available at: https://github.com/WuXuanxin/human-eval-llm-simplification
Information Retrieval 11
☆ Item-centric Exploration for Cold Start Problem
Recommender systems face a critical challenge in the item cold-start problem, which limits content diversity and exacerbates popularity bias by struggling to recommend new items. While existing solutions often rely on auxiliary data, but this paper illuminates a distinct, yet equally pressing, issue stemming from the inherent user-centricity of many recommender systems. We argue that in environments with large and rapidly expanding item inventories, the traditional focus on finding the "best item for a user" can inadvertently obscure the ideal audience for nascent content. To counter this, we introduce the concept of item-centric recommendations, shifting the paradigm to identify the optimal users for new items. Our initial realization of this vision involves an item-centric control integrated into an exploration system. This control employs a Bayesian model with Beta distributions to assess candidate items based on a predicted balance between user satisfaction and the item's inherent quality. Empirical online evaluations reveal that this straightforward control markedly improves cold-start targeting efficacy, enhances user satisfaction with newly explored content, and significantly increases overall exploration efficiency.
comment: Accepted for publication on 2025 ACM Recsys Conference Industry Track
☆ Balancing Semantic Relevance and Engagement in Related Video Recommendations
Related video recommendations commonly use collaborative filtering (CF) driven by co-engagement signals, often resulting in recommendations lacking semantic coherence and exhibiting strong popularity bias. This paper introduces a novel multi-objective retrieval framework, enhancing standard two-tower models to explicitly balance semantic relevance and user engagement. Our approach uniquely combines: (a) multi-task learning (MTL) to jointly optimize co-engagement and semantic relevance, explicitly prioritizing topical coherence; (b) fusion of multimodal content features (textual and visual embeddings) for richer semantic understanding; and (c) off-policy correction (OPC) via inverse propensity weighting to effectively mitigate popularity bias. Evaluation on industrial-scale data and a two-week live A/B test reveals our framework's efficacy. We observed significant improvements in semantic relevance (from 51% to 63% topic match rate), a reduction in popular item distribution (-13.8% popular video recommendations), and a +0.04% improvement in our topline user engagement metric. Our method successfully achieves better semantic coherence, balanced engagement, and practical scalability for real-world deployment.
☆ Knowledge Conceptualization Impacts RAG Efficacy
Explainability and interpretability are cornerstones of frontier and next-generation artificial intelligence (AI) systems. This is especially true in recent systems, such as large language models (LLMs), and more broadly, generative AI. On the other hand, adaptability to new domains, contexts, or scenarios is also an important aspect for a successful system. As such, we are particularly interested in how we can merge these two efforts, that is, investigating the design of transferable and interpretable neurosymbolic AI systems. Specifically, we focus on a class of systems referred to as ''Agentic Retrieval-Augmented Generation'' systems, which actively select, interpret, and query knowledge sources in response to natural language prompts. In this paper, we systematically evaluate how different conceptualizations and representations of knowledge, particularly the structure and complexity, impact an AI agent (in this case, an LLM) in effectively querying a triplestore. We report our results, which show that there are impacts from both approaches, and we discuss their impact and implications.
☆ Correcting the LogQ Correction: Revisiting Sampled Softmax for Large-Scale Retrieval RecSys 2025
Two-tower neural networks are a popular architecture for the retrieval stage in recommender systems. These models are typically trained with a softmax loss over the item catalog. However, in web-scale settings, the item catalog is often prohibitively large, making full softmax infeasible. A common solution is sampled softmax, which approximates the full softmax using a small number of sampled negatives. One practical and widely adopted approach is to use in-batch negatives, where negatives are drawn from items in the current mini-batch. However, this introduces a bias: items that appear more frequently in the batch (i.e., popular items) are penalized more heavily. To mitigate this issue, a popular industry technique known as logQ correction adjusts the logits during training by subtracting the log-probability of an item appearing in the batch. This correction is derived by analyzing the bias in the gradient and applying importance sampling, effectively twice, using the in-batch distribution as a proposal distribution. While this approach improves model quality, it does not fully eliminate the bias. In this work, we revisit the derivation of logQ correction and show that it overlooks a subtle but important detail: the positive item in the denominator is not Monte Carlo-sampled - it is always present with probability 1. We propose a refined correction formula that accounts for this. Notably, our loss introduces an interpretable sample weight that reflects the model's uncertainty - the probability of misclassification under the current parameters. We evaluate our method on both public and proprietary datasets, demonstrating consistent improvements over the standard logQ correction.
comment: Accepted at ACM RecSys 2025. Author's version. To appear in the Proceedings of the 18th ACM Conference on Recommender Systems
☆ Ambiguity-Aware and High-Order Relation Learning for Multi-Grained Image-Text Matching
Image-text matching is crucial for bridging the semantic gap between computer vision and natural language processing. However, existing methods still face challenges in handling high-order associations and semantic ambiguities among similar instances. These ambiguities arise from subtle differences between soft positive samples (semantically similar but incorrectly labeled) and soft negative samples (locally matched but globally inconsistent), creating matching uncertainties. Furthermore, current methods fail to fully utilize the neighborhood relationships among semantically similar instances within training batches, limiting the model's ability to learn high-order shared knowledge. This paper proposes the Ambiguity-Aware and High-order Relation learning framework (AAHR) to address these issues. AAHR constructs a unified representation space through dynamic clustering prototype contrastive learning, effectively mitigating the soft positive sample problem. The framework introduces global and local feature extraction mechanisms and an adaptive aggregation network, significantly enhancing full-grained semantic understanding capabilities. Additionally, AAHR employs intra-modal and inter-modal correlation matrices to investigate neighborhood relationships among sample instances thoroughly. It incorporates GNN to enhance semantic interactions between instances. Furthermore, AAHR integrates momentum contrastive learning to expand the negative sample set. These combined strategies significantly improve the model's ability to discriminate between features. Experimental results demonstrate that AAHR outperforms existing state-of-the-art methods on Flickr30K, MSCOCO, and ECCV Caption datasets, considerably improving the accuracy and efficiency of image-text matching. The code and model checkpoints for this research are available at https://github.com/Image-Text-Matching/AAHR .
comment: Accepted by the Knowledge-Based Systems(KBS), 2025
☆ Retrieval-Augmented Recommendation Explanation Generation with Hierarchical Aggregation
Explainable Recommender System (ExRec) provides transparency to the recommendation process, increasing users' trust and boosting the operation of online services. With the rise of large language models (LLMs), whose extensive world knowledge and nuanced language understanding enable the generation of human-like, contextually grounded explanations, LLM-powered ExRec has gained great momentum. However, existing LLM-based ExRec models suffer from profile deviation and high retrieval overhead, hindering their deployment. To address these issues, we propose Retrieval-Augmented Recommendation Explanation Generation with Hierarchical Aggregation (REXHA). Specifically, we design a hierarchical aggregation based profiling module that comprehensively considers user and item review information, hierarchically summarizing and constructing holistic profiles. Furthermore, we introduce an efficient retrieval module using two types of pseudo-document queries to retrieve relevant reviews to enhance the generation of recommendation explanations, effectively reducing retrieval latency and improving the recall of relevant reviews. Extensive experiments demonstrate that our method outperforms existing approaches by up to 12.6% w.r.t. the explanation quality while achieving high retrieval efficiency.
☆ DS@GT at Touché: Large Language Models for Retrieval-Augmented Debate
Large Language Models (LLMs) demonstrate strong conversational abilities. In this Working Paper, we study them in the context of debating in two ways: their ability to perform in a structured debate along with a dataset of arguments to use and their ability to evaluate utterances throughout the debate. We deploy six leading publicly available models from three providers for the Retrieval-Augmented Debate and Evaluation. The evaluation is performed by measuring four key metrics: Quality, Quantity, Manner, and Relation. Throughout this task, we found that although LLMs perform well in debates when given related arguments, they tend to be verbose in responses yet consistent in evaluation. The accompanying source code for this paper is located at https://github.com/dsgt-arc/touche-2025-rad.
♻ ☆ Towards Two-Stage Counterfactual Learning to Rank ICTIR 2025
Counterfactual learning to rank (CLTR) aims to learn a ranking policy from user interactions while correcting for the inherent biases in interaction data, such as position bias. Existing CLTR methods assume a single ranking policy that selects top-K ranking from the entire document candidate set. In real-world applications, the candidate document set is on the order of millions, making a single-stage ranking policy impractical. In order to scale to millions of documents, real-world ranking systems are designed in a two-stage fashion, with a candidate generator followed by a ranker. The existing CLTR method for a two-stage offline ranking system only considers the top-1 ranking set-up and only focuses on training the candidate generator, with the ranker fixed. A CLTR method for training both the ranker and candidate generator jointly is missing from the existing literature. In this paper, we propose a two-stage CLTR estimator that considers the interaction between the two stages and estimates the joint value of the two policies offline. In addition, we propose a novel joint optimization method to train the candidate and ranker policies, respectively. To the best of our knowledge, we are the first to propose a CLTR estimator and learning method for two-stage ranking. Experimental results on a semi-synthetic benchmark demonstrate the effectiveness of the proposed joint CLTR method over baselines.
comment: Accepted at ICTIR 2025 (co-located with SIGIR 2025)
♻ ☆ DTECT: Dynamic Topic Explorer & Context Tracker
The explosive growth of textual data over time presents a significant challenge in uncovering evolving themes and trends. Existing dynamic topic modeling techniques, while powerful, often exist in fragmented pipelines that lack robust support for interpretation and user-friendly exploration. We introduce DTECT (Dynamic Topic Explorer & Context Tracker), an end-to-end system that bridges the gap between raw textual data and meaningful temporal insights. DTECT provides a unified workflow that supports data preprocessing, multiple model architectures, and dedicated evaluation metrics to analyze the topic quality of temporal topic models. It significantly enhances interpretability by introducing LLM-driven automatic topic labeling, trend analysis via temporally salient words, interactive visualizations with document-level summarization, and a natural language chat interface for intuitive data querying. By integrating these features into a single, cohesive platform, DTECT empowers users to more effectively track and understand thematic dynamics. DTECT is open-source and available at https://github.com/AdhyaSuman/DTECT.
comment: Code: https://github.com/AdhyaSuman/DTECT | Demo: https://huggingface.co/spaces/AdhyaSuman/DTECT | Video: https://youtu.be/B8nNfxFoJAU
♻ ☆ SymRAG: Efficient Neuro-Symbolic Retrieval Through Adaptive Query Routing
Current Retrieval-Augmented Generation systems use uniform processing, causing inefficiency as simple queries consume resources similar to complex multi-hop tasks. We present SymRAG, a framework that introduces adaptive query routing via real-time complexity and load assessment to select symbolic, neural, or hybrid pathways. SymRAG's neuro-symbolic approach adjusts computational pathways based on both query characteristics and system load, enabling efficient resource allocation across diverse query types. By combining linguistic and structural query properties with system load metrics, SymRAG allocates resources proportional to reasoning requirements. Evaluated on 2,000 queries across HotpotQA (multi-hop reasoning) and DROP (discrete reasoning) using Llama-3.2-3B and Mistral-7B models, SymRAG achieves competitive accuracy (97.6--100.0% exact match) with efficient resource utilization (3.6--6.2% CPU utilization, 0.985--3.165s processing). Disabling adaptive routing increases processing time by 169--1151%, showing its significance for complex models. These results suggest adaptive computation strategies are more sustainable and scalable for hybrid AI systems that use dynamic routing and neuro-symbolic frameworks.
comment: Accepted at 19th International Conference on Neurosymbolic Learning and Reasoning (NeSy 2025)
♻ ☆ RadIR: A Scalable Framework for Multi-Grained Medical Image Retrieval via Radiology Report Mining
Developing advanced medical imaging retrieval systems is challenging due to the varying definitions of `similar images' across different medical contexts. This challenge is compounded by the lack of large-scale, high-quality medical imaging retrieval datasets and benchmarks. In this paper, we propose a novel methodology that leverages dense radiology reports to define image-wise similarity ordering at multiple granularities in a scalable and fully automatic manner. Using this approach, we construct two comprehensive medical imaging retrieval datasets: MIMIC-IR for Chest X-rays and CTRATE-IR for CT scans, providing detailed image-image ranking annotations conditioned on diverse anatomical structures. Furthermore, we develop two retrieval systems, RadIR-CXR and model-ChestCT, which demonstrate superior performance in traditional image-image and image-report retrieval tasks. These systems also enable flexible, effective image retrieval conditioned on specific anatomical structures described in text, achieving state-of-the-art results on 77 out of 78 metrics.
Multimedia 4
☆ Balancing Semantic Relevance and Engagement in Related Video Recommendations
Related video recommendations commonly use collaborative filtering (CF) driven by co-engagement signals, often resulting in recommendations lacking semantic coherence and exhibiting strong popularity bias. This paper introduces a novel multi-objective retrieval framework, enhancing standard two-tower models to explicitly balance semantic relevance and user engagement. Our approach uniquely combines: (a) multi-task learning (MTL) to jointly optimize co-engagement and semantic relevance, explicitly prioritizing topical coherence; (b) fusion of multimodal content features (textual and visual embeddings) for richer semantic understanding; and (c) off-policy correction (OPC) via inverse propensity weighting to effectively mitigate popularity bias. Evaluation on industrial-scale data and a two-week live A/B test reveals our framework's efficacy. We observed significant improvements in semantic relevance (from 51% to 63% topic match rate), a reduction in popular item distribution (-13.8% popular video recommendations), and a +0.04% improvement in our topline user engagement metric. Our method successfully achieves better semantic coherence, balanced engagement, and practical scalability for real-world deployment.
☆ Acoustic Wave Modeling Using 2D FDTD: Applications in Unreal Engine For Dynamic Sound Rendering
Accurate sound propagation simulation is essential for delivering immersive experiences in virtual applications, yet industry methods for acoustic modeling often do not account for the full breadth of acoustic wave phenomena. This paper proposes a novel two-dimensional (2D) finite-difference time-domain (FDTD) framework that simulates sound propagation as a wave-based model in Unreal Engine, with an emphasis on capturing lower frequency wave phenomena, embedding occlusion, diffraction, reflection and interference in generated impulse responses. The process begins by discretizing the scene geometry into a 2D grid via a top-down projection from which obstacle masks and boundary conditions are derived. A Python-based FDTD solver injects a sine sweep at a source position, and virtual quadraphonic microphone arrays record pressure field responses at pre-defined listener positions. De-convolution of the pressure responses yields multi-channel impulse responses that retain spatial directionality which are then integrated into Unreal Engine's audio pipeline for dynamic playback. Benchmark tests confirm agreement with analytical expectations, and the paper outlines hybrid extensions aimed at commercial viability.
comment: Accepted to the 50th International Computer Music Conference (ICMC), 2025
☆ Ambiguity-Aware and High-Order Relation Learning for Multi-Grained Image-Text Matching
Image-text matching is crucial for bridging the semantic gap between computer vision and natural language processing. However, existing methods still face challenges in handling high-order associations and semantic ambiguities among similar instances. These ambiguities arise from subtle differences between soft positive samples (semantically similar but incorrectly labeled) and soft negative samples (locally matched but globally inconsistent), creating matching uncertainties. Furthermore, current methods fail to fully utilize the neighborhood relationships among semantically similar instances within training batches, limiting the model's ability to learn high-order shared knowledge. This paper proposes the Ambiguity-Aware and High-order Relation learning framework (AAHR) to address these issues. AAHR constructs a unified representation space through dynamic clustering prototype contrastive learning, effectively mitigating the soft positive sample problem. The framework introduces global and local feature extraction mechanisms and an adaptive aggregation network, significantly enhancing full-grained semantic understanding capabilities. Additionally, AAHR employs intra-modal and inter-modal correlation matrices to investigate neighborhood relationships among sample instances thoroughly. It incorporates GNN to enhance semantic interactions between instances. Furthermore, AAHR integrates momentum contrastive learning to expand the negative sample set. These combined strategies significantly improve the model's ability to discriminate between features. Experimental results demonstrate that AAHR outperforms existing state-of-the-art methods on Flickr30K, MSCOCO, and ECCV Caption datasets, considerably improving the accuracy and efficiency of image-text matching. The code and model checkpoints for this research are available at https://github.com/Image-Text-Matching/AAHR .
comment: Accepted by the Knowledge-Based Systems(KBS), 2025
♻ ☆ MSVD-Indonesian: A Benchmark for Multimodal Video-Text Tasks in Indonesian
Multimodal learning on video and text has seen significant progress, particularly in tasks like text-to-video retrieval, video-to-text retrieval, and video captioning. However, most existing methods and datasets focus exclusively on English. Despite Indonesian being one of the most widely spoken languages, multimodal research in Indonesian remains under-explored, largely due to the lack of benchmark datasets. To address this gap, we introduce the first public Indonesian video-text dataset by translating the English captions in the MSVD dataset into Indonesian. Using this dataset, we evaluate neural network models which were developed for the English video-text dataset on three tasks, i.e., text-to-video retrieval, video-to-text retrieval, and video captioning. Most existing models rely on feature extractors pretrained on English vision-language datasets, raising concerns about their applicability to Indonesian, given the scarcity of large-scale pretraining resources in the language. We apply a cross-lingual transfer learning approach by leveraging English-pretrained extractors and fine-tuning models on our Indonesian dataset. Experimental results demonstrate that this strategy improves performance across all tasks and metrics. We release our dataset publicly to support future research and hope it will inspire further progress in Indonesian multimodal learning.
comment: 10 pages, 5 figures, 5 tables
Computation and Language 72
☆ NeuralOS: Towards Simulating Operating Systems via Neural Generative Models
We introduce NeuralOS, a neural framework that simulates graphical user interfaces (GUIs) of operating systems by directly predicting screen frames in response to user inputs such as mouse movements, clicks, and keyboard events. NeuralOS combines a recurrent neural network (RNN), which tracks computer state, with a diffusion-based neural renderer that generates screen images. The model is trained on a large-scale dataset of Ubuntu XFCE recordings, which include both randomly generated interactions and realistic interactions produced by AI agents. Experiments show that NeuralOS successfully renders realistic GUI sequences, accurately captures mouse interactions, and reliably predicts state transitions like application launches. Although modeling fine-grained keyboard interactions precisely remains challenging, NeuralOS offers a step toward creating fully adaptive, generative neural interfaces for future human-computer interaction systems.
☆ KV Cache Steering for Inducing Reasoning in Small Language Models
We propose cache steering, a lightweight method for implicit steering of language models via a one-shot intervention applied directly to the key-value cache. To validate its effectiveness, we apply cache steering to induce chain-of-thought reasoning in small language models. Our approach leverages GPT-4o-generated reasoning traces to construct steering vectors that shift model behavior toward more explicit, multi-step reasoning without fine-tuning or prompt modifications. Experimental evaluations on diverse reasoning benchmarks demonstrate that cache steering improves both the qualitative structure of model reasoning and quantitative task performance. Compared to prior activation steering techniques that require continuous interventions, our one-shot cache steering offers substantial advantages in terms of hyperparameter stability, inference-time efficiency, and ease of integration, making it a more robust and practical solution for controlled generation.
☆ One Token to Fool LLM-as-a-Judge
Generative reward models (also known as LLMs-as-judges), which use large language models (LLMs) to evaluate answer quality, are increasingly adopted in reinforcement learning with verifiable rewards (RLVR). They are often preferred over rigid rule-based metrics, especially for complex reasoning tasks involving free-form outputs. In this paradigm, an LLM is typically prompted to compare a candidate answer against a ground-truth reference and assign a binary reward indicating correctness. Despite the seeming simplicity of this comparison task, we find that generative reward models exhibit surprising vulnerabilities to superficial manipulations: non-word symbols (e.g., ":" or ".") or reasoning openers like "Thought process:" and "Let's solve this problem step by step." can often lead to false positive rewards. We demonstrate that this weakness is widespread across LLMs, datasets, and prompt formats, posing a serious threat for core algorithmic paradigms that rely on generative reward models, such as rejection sampling, preference optimization, and RLVR. To mitigate this issue, we introduce a simple yet effective data augmentation strategy and train a new generative reward model with substantially improved robustness. Our findings highlight the urgent need for more reliable LLM-based evaluation methods. We release our robust, general-domain reward model and its synthetic training data at https://huggingface.co/sarosavo/Master-RM and https://huggingface.co/datasets/sarosavo/Master-RM.
☆ BlockFFN: Towards End-Side Acceleration-Friendly Mixture-of-Experts with Chunk-Level Activation Sparsity
To alleviate the computational burden of large language models (LLMs), architectures with activation sparsity, represented by mixture-of-experts (MoE), have attracted increasing attention. However, the non-differentiable and inflexible routing of vanilla MoE hurts model performance. Moreover, while each token activates only a few parameters, these sparsely-activated architectures exhibit low chunk-level sparsity, indicating that the union of multiple consecutive tokens activates a large ratio of parameters. Such a sparsity pattern is unfriendly for acceleration under low-resource conditions (e.g., end-side devices) and incompatible with mainstream acceleration techniques (e.g., speculative decoding). To address these challenges, we introduce a novel MoE architecture, BlockFFN, as well as its efficient training and deployment techniques. Specifically, we use a router integrating ReLU activation and RMSNorm for differentiable and flexible routing. Next, to promote both token-level sparsity (TLS) and chunk-level sparsity (CLS), CLS-aware training objectives are designed, making BlockFFN more acceleration-friendly. Finally, we implement efficient acceleration kernels, combining activation sparsity and speculative decoding for the first time. The experimental results demonstrate the superior performance of BlockFFN over other MoE baselines, achieving over 80% TLS and 70% 8-token CLS. Our kernels achieve up to 3.67$\times$ speedup on real end-side devices than dense models. All codes and checkpoints are available publicly (https://github.com/thunlp/BlockFFN).
comment: 21 pages, 7 figures, 15 tables
☆ On Barriers to Archival Audio Processing
In this study, we leverage a unique UNESCO collection of mid-20th century radio recordings to probe the robustness of modern off-the-shelf language identification (LID) and speaker recognition (SR) methods, especially with respect to the impact of multilingual speakers and cross-age recordings. Our findings suggest that LID systems, such as Whisper, are increasingly adept at handling second-language and accented speech. However, speaker embeddings remain a fragile component of speech processing pipelines that is prone to biases related to the channel, age, and language. Issues which will need to be overcome should archives aim to employ SR methods for speaker indexing.
comment: Update with Acknowledgements of ICNSLP 2025 paper
☆ Multilingual Multimodal Software Developer for Code Generation
The rapid advancement of Large Language Models (LLMs) has significantly improved code generation, yet most models remain text-only, neglecting crucial visual aids like diagrams and flowcharts used in real-world software development. To bridge this gap, we introduce MM-Coder, a Multilingual Multimodal software developer. MM-Coder integrates visual design inputs-Unified Modeling Language (UML) diagrams and flowcharts (termed Visual Workflow)-with textual instructions to enhance code generation accuracy and architectural alignment. To enable this, we developed MMc-Instruct, a diverse multimodal instruction-tuning dataset including visual-workflow-based code generation, allowing MM-Coder to synthesize textual and graphical information like human developers, distinct from prior work on narrow tasks. Furthermore, we introduce MMEval, a new benchmark for evaluating multimodal code generation, addressing existing text-only limitations. Our evaluations using MMEval highlight significant remaining challenges for models in precise visual information capture, instruction following, and advanced programming knowledge. Our work aims to revolutionize industrial programming by enabling LLMs to interpret and implement complex specifications conveyed through both text and visual designs.
comment: Preprint
☆ KG-Attention: Knowledge Graph-Guided Attention at Test-Time via Bidirectional Information Aggregation
Knowledge graphs (KGs) play a critical role in enhancing large language models (LLMs) by introducing structured and grounded knowledge into the learning process. However, most existing KG-enhanced approaches rely on parameter-intensive fine-tuning, which risks catastrophic forgetting and degrades the pretrained model's generalization. Moreover, they exhibit limited adaptability to real-time knowledge updates due to their static integration frameworks. To address these issues, we introduce the first test-time KG-augmented framework for LLMs, built around a dedicated knowledge graph-guided attention (KGA) module that enables dynamic knowledge fusion without any parameter updates. The proposed KGA module augments the standard self-attention mechanism with two synergistic pathways: outward and inward aggregation. Specifically, the outward pathway dynamically integrates external knowledge into input representations via input-driven KG fusion. This inward aggregation complements the outward pathway by refining input representations through KG-guided filtering, suppressing task-irrelevant signals and amplifying knowledge-relevant patterns. Importantly, while the outward pathway handles knowledge fusion, the inward path selects the most relevant triples and feeds them back into the fusion process, forming a closed-loop enhancement mechanism. By synergistically combining these two pathways, the proposed method supports real-time knowledge fusion exclusively at test-time, without any parameter modification. Extensive experiments on five benchmarks verify the comparable knowledge fusion performance of KGA.
☆ KELPS: A Framework for Verified Multi-Language Autoformalization via Semantic-Syntactic Alignment ICML 2025
Modern large language models (LLMs) show promising progress in formalizing informal mathematics into machine-verifiable theorems. However, these methods still face bottlenecks due to the limited quantity and quality of multilingual parallel corpora. In this paper, we propose a novel neuro-symbolic framework KELPS (Knowledge-Equation based Logical Processing System) to address these problems. KELPS is an iterative framework for translating, synthesizing, and filtering informal data into multiple formal languages (Lean, Coq, and Isabelle). First, we translate natural language into Knowledge Equations (KEs), a novel language that we designed, theoretically grounded in assertional logic. Next, we convert them to target languages through rigorously defined rules that preserve both syntactic structure and semantic meaning. This process yielded a parallel corpus of over 60,000 problems. Our framework achieves 88.9% syntactic accuracy (pass@1) on MiniF2F, outperforming SOTA models such as Deepseek-V3 (81%) and Herald (81.3%) across multiple datasets. All datasets and codes are available in the supplementary materials.
comment: Accepted by the ICML 2025 AI4MATH Workshop. 22 pages, 16 figures, 2 tables
☆ The Impact of Automatic Speech Transcription on Speaker Attribution
Speaker attribution from speech transcripts is the task of identifying a speaker from the transcript of their speech based on patterns in their language use. This task is especially useful when the audio is unavailable (e.g. deleted) or unreliable (e.g. anonymized speech). Prior work in this area has primarily focused on the feasibility of attributing speakers using transcripts produced by human annotators. However, in real-world settings, one often only has more errorful transcripts produced by automatic speech recognition (ASR) systems. In this paper, we conduct what is, to our knowledge, the first comprehensive study of the impact of automatic transcription on speaker attribution performance. In particular, we study the extent to which speaker attribution performance degrades in the face of transcription errors, as well as how properties of the ASR system impact attribution. We find that attribution is surprisingly resilient to word-level transcription errors and that the objective of recovering the true transcript is minimally correlated with attribution performance. Overall, our findings suggest that speaker attribution on more errorful transcripts produced by ASR is as good, if not better, than attribution based on human-transcribed data, possibly because ASR transcription errors can capture speaker-specific features revealing of speaker identity.
☆ Scaling Attention to Very Long Sequences in Linear Time with Wavelet-Enhanced Random Spectral Attention (WERSA)
Transformer models are computationally costly on long sequences since regular attention has quadratic $O(n^2)$ time complexity. We introduce Wavelet-Enhanced Random Spectral Attention (WERSA), a novel mechanism of linear $O(n)$ time complexity that is pivotal to enable successful long-sequence processing without the performance trade-off. WERSA merges content-adaptive random spectral features together with multi-resolution Haar wavelets and learnable parameters to selectively attend to informative scales of data while preserving linear efficiency. Large-scale comparisons \textbf{on single GPU} and across various benchmarks (vision, NLP, hierarchical reasoning) and various attention mechanisms (like Multiheaded Attention, Flash-Attention-2, FNet, Linformer, Performer, Waveformer), reveal uniform advantages of WERSA. It achieves best accuracy in all tests. On ArXiv classification, WERSA improves accuracy over vanilla attention by 1.2\% (86.2\% vs 85.0\%) while cutting training time by 81\% (296s vs 1554s) and FLOPS by 73.4\% (26.2G vs 98.4G). Significantly, WERSA excels where vanilla and FlashAttention-2 fail: on ArXiv-128k's extremely lengthy sequences, it achieves best accuracy (79.1\%) and AUC (0.979) among viable methods, operating on data that gives Out-Of-Memory errors to quadratic methods while being \textbf{twice as fast} as Waveformer, its next-best competitor. By significantly reducing computational loads without compromising accuracy, WERSA makes possible more practical, more affordable, long-context models, in particular on low-resource hardware, for more sustainable and more scalable AI development.
comment: 10 pages, 1 figure
☆ A comprehensive study of LLM-based argument classification: from LLAMA through GPT-4o to Deepseek-R1
Argument mining (AM) is an interdisciplinary research field that integrates insights from logic, philosophy, linguistics, rhetoric, law, psychology, and computer science. It involves the automatic identification and extraction of argumentative components, such as premises and claims, and the detection of relationships between them, such as support, attack, or neutrality. Recently, the field has advanced significantly, especially with the advent of large language models (LLMs), which have enhanced the efficiency of analyzing and extracting argument semantics compared to traditional methods and other deep learning models. There are many benchmarks for testing and verifying the quality of LLM, but there is still a lack of research and results on the operation of these models in publicly available argument classification databases. This paper presents a study of a selection of LLM's, using diverse datasets such as Args.me and UKP. The models tested include versions of GPT, Llama, and DeepSeek, along with reasoning-enhanced variants incorporating the Chain-of-Thoughts algorithm. The results indicate that ChatGPT-4o outperforms the others in the argument classification benchmarks. In case of models incorporated with reasoning capabilities, the Deepseek-R1 shows its superiority. However, despite their superiority, GPT-4o and Deepseek-R1 still make errors. The most common errors are discussed for all models. To our knowledge, the presented work is the first broader analysis of the mentioned datasets using LLM and prompt algorithms. The work also shows some weaknesses of known prompt algorithms in argument analysis, while indicating directions for their improvement. The added value of the work is the in-depth analysis of the available argument datasets and the demonstration of their shortcomings.
☆ DocPolarBERT: A Pre-trained Model for Document Understanding with Relative Polar Coordinate Encoding of Layout Structures
We introduce DocPolarBERT, a layout-aware BERT model for document understanding that eliminates the need for absolute 2D positional embeddings. We extend self-attention to take into account text block positions in relative polar coordinate system rather than the Cartesian one. Despite being pre-trained on a dataset more than six times smaller than the widely used IIT-CDIP corpus, DocPolarBERT achieves state-of-the-art results. These results demonstrate that a carefully designed attention mechanism can compensate for reduced pre-training data, offering an efficient and effective alternative for document understanding.
☆ Large Multi-modal Model Cartographic Map Comprehension for Textual Locality Georeferencing
Millions of biological sample records collected in the last few centuries archived in natural history collections are un-georeferenced. Georeferencing complex locality descriptions associated with these collection samples is a highly labour-intensive task collection agencies struggle with. None of the existing automated methods exploit maps that are an essential tool for georeferencing complex relations. We present preliminary experiments and results of a novel method that exploits multi-modal capabilities of recent Large Multi-Modal Models (LMM). This method enables the model to visually contextualize spatial relations it reads in the locality description. We use a grid-based approach to adapt these auto-regressive models for this task in a zero-shot setting. Our experiments conducted on a small manually annotated dataset show impressive results for our approach ($\sim$1 km Average distance error) compared to uni-modal georeferencing with Large Language Models and existing georeferencing tools. The paper also discusses the findings of the experiments in light of an LMM's ability to comprehend fine-grained maps. Motivated by these results, a practical framework is proposed to integrate this method into a georeferencing workflow.
☆ The AI Language Proficiency Monitor -- Tracking the Progress of LLMs on Multilingual Benchmarks
To ensure equitable access to the benefits of large language models (LLMs), it is essential to evaluate their capabilities across the world's languages. We introduce the AI Language Proficiency Monitor, a comprehensive multilingual benchmark that systematically assesses LLM performance across up to 200 languages, with a particular focus on low-resource languages. Our benchmark aggregates diverse tasks including translation, question answering, math, and reasoning, using datasets such as FLORES+, MMLU, GSM8K, TruthfulQA, and ARC. We provide an open-source, auto-updating leaderboard and dashboard that supports researchers, developers, and policymakers in identifying strengths and gaps in model performance. In addition to ranking models, the platform offers descriptive insights such as a global proficiency map and trends over time. By complementing and extending prior multilingual benchmarks, our work aims to foster transparency, inclusivity, and progress in multilingual AI. The system is available at https://huggingface.co/spaces/fair-forward/evals-for-every-language.
☆ A Multi-granularity Concept Sparse Activation and Hierarchical Knowledge Graph Fusion Framework for Rare Disease Diagnosis
Despite advances from medical large language models in healthcare, rare-disease diagnosis remains hampered by insufficient knowledge-representation depth, limited concept understanding, and constrained clinical reasoning. We propose a framework that couples multi-granularity sparse activation of medical concepts with a hierarchical knowledge graph. Four complementary matching algorithms, diversity control, and a five-level fallback strategy enable precise concept activation, while a three-layer knowledge graph (taxonomy, clinical features, instances) provides structured, up-to-date context. Experiments on the BioASQ rare-disease QA set show BLEU gains of 0.09, ROUGE gains of 0.05, and accuracy gains of 0.12, with peak accuracy of 0.89 approaching the 0.90 clinical threshold. Expert evaluation confirms improvements in information quality, reasoning, and professional expression, suggesting our approach shortens the "diagnostic odyssey" for rare-disease patients.
comment: 10 pages,3 figures
☆ PromotionGo at SemEval-2025 Task 11: A Feature-Centric Framework for Cross-Lingual Multi-Emotion Detection in Short Texts
This paper presents our system for SemEval 2025 Task 11: Bridging the Gap in Text-Based Emotion Detection (Track A), which focuses on multi-label emotion detection in short texts. We propose a feature-centric framework that dynamically adapts document representations and learning algorithms to optimize language-specific performance. Our study evaluates three key components: document representation, dimensionality reduction, and model training in 28 languages, highlighting five for detailed analysis. The results show that TF-IDF remains highly effective for low-resource languages, while contextual embeddings like FastText and transformer-based document representations, such as those produced by Sentence-BERT, exhibit language-specific strengths. Principal Component Analysis (PCA) reduces training time without compromising performance, particularly benefiting FastText and neural models such as Multi-Layer Perceptrons (MLP). Computational efficiency analysis underscores the trade-off between model complexity and processing cost. Our framework provides a scalable solution for multilingual emotion detection, addressing the challenges of linguistic diversity and resource constraints.
☆ Semantic-Augmented Latent Topic Modeling with LLM-in-the-Loop
Latent Dirichlet Allocation (LDA) is a prominent generative probabilistic model used for uncovering abstract topics within document collections. In this paper, we explore the effectiveness of augmenting topic models with Large Language Models (LLMs) through integration into two key phases: Initialization and Post-Correction. Since the LDA is highly dependent on the quality of its initialization, we conduct extensive experiments on the LLM-guided topic clustering for initializing the Gibbs sampling algorithm. Interestingly, the experimental results reveal that while the proposed initialization strategy improves the early iterations of LDA, it has no effect on the convergence and yields the worst performance compared to the baselines. The LLM-enabled post-correction, on the other hand, achieved a promising improvement of 5.86% in the coherence evaluation. These results highlight the practical benefits of the LLM-in-the-loop approach and challenge the belief that LLMs are always the superior text mining alternative.
☆ LLaPa: A Vision-Language Model Framework for Counterfactual-Aware Procedural Planning
While large language models (LLMs) have advanced procedural planning for embodied AI systems through strong reasoning abilities, the integration of multimodal inputs and counterfactual reasoning remains underexplored. To tackle these challenges, we introduce LLaPa, a vision-language model framework designed for multimodal procedural planning. LLaPa generates executable action sequences from textual task descriptions and visual environmental images using vision-language models (VLMs). Furthermore, we enhance LLaPa with two auxiliary modules to improve procedural planning. The first module, the Task-Environment Reranker (TER), leverages task-oriented segmentation to create a task-sensitive feature space, aligning textual descriptions with visual environments and emphasizing critical regions for procedural execution. The second module, the Counterfactual Activities Retriever (CAR), identifies and emphasizes potential counterfactual conditions, enhancing the model's reasoning capability in counterfactual scenarios. Extensive experiments on ActPlan-1K and ALFRED benchmarks demonstrate that LLaPa generates higher-quality plans with superior LCS and correctness, outperforming advanced models. The code and models are available https://github.com/sunshibo1234/LLaPa.
☆ A Third Paradigm for LLM Evaluation: Dialogue Game-Based Evaluation using clembench
There are currently two main paradigms for evaluating large language models (LLMs), reference-based evaluation and preference-based evaluation. The first, carried over from the evaluation of machine learning models in general, relies on pre-defined task instances, for which reference task executions are available. The second, best exemplified by the LM-arena, relies on (often self-selected) users bringing their own intents to a site that routes these to several models in parallel, among whose responses the user then selects their most preferred one. The former paradigm hence excels at control over what is tested, while the latter comes with higher ecological validity, testing actual use cases interactively. Recently, a third complementary paradigm has emerged that combines some of the strengths of these approaches, offering control over multi-turn, reference-free, repeatable interactions, while stressing goal-directedness: dialogue game based evaluation. While the utility of this approach has been shown by several projects, its adoption has been held back by the lack of a mature, easily re-usable implementation. In this paper, we present clembench, which has been in continuous development since 2023 and has in its latest release been optimized for ease of general use. We describe how it can be used to benchmark one's own models (using a provided set of benchmark game instances in English), as well as how easily the benchmark itself can be extended with new, tailor-made targeted tests.
comment: All code required to run the benchmark, as well as extensive documentation, is available at https://github.com/clembench/clembench
☆ Enhancing Essay Cohesion Assessment: A Novel Item Response Theory Approach
Essays are considered a valuable mechanism for evaluating learning outcomes in writing. Textual cohesion is an essential characteristic of a text, as it facilitates the establishment of meaning between its parts. Automatically scoring cohesion in essays presents a challenge in the field of educational artificial intelligence. The machine learning algorithms used to evaluate texts generally do not consider the individual characteristics of the instances that comprise the analysed corpus. In this meaning, item response theory can be adapted to the context of machine learning, characterising the ability, difficulty and discrimination of the models used. This work proposes and analyses the performance of a cohesion score prediction approach based on item response theory to adjust the scores generated by machine learning models. In this study, the corpus selected for the experiments consisted of the extended Essay-BR, which includes 6,563 essays in the style of the National High School Exam (ENEM), and the Brazilian Portuguese Narrative Essays, comprising 1,235 essays written by 5th to 9th grade students from public schools. We extracted 325 linguistic features and treated the problem as a machine learning regression task. The experimental results indicate that the proposed approach outperforms conventional machine learning models and ensemble methods in several evaluation metrics. This research explores a potential approach for improving the automatic evaluation of cohesion in educational essays.
comment: 24 pages, 4 tables
☆ ILT-Iterative LoRA Training through Focus-Feedback-Fix for Multilingual Speech Recognition
The deep integration of large language models and automatic speech recognition systems has become a promising research direction with high practical value. To address the overfitting issue commonly observed in Low-Rank Adaptation (LoRA) during the supervised fine-tuning (SFT) stage, this work proposes an innovative training paradigm Iterative LoRA Training (ILT) in combination with an Iterative Pseudo Labeling strategy, effectively enhancing the theoretical upper bound of model performance. Based on Whisper-large-v3 and Qwen2-Audio, we conduct systematic experiments using a three-stage training process: Focus Training, Feed Back Training, and Fix Training. Experimental results demonstrate the effectiveness of the proposed method. Furthermore, the MegaAIS research team applied this technique in the Interspeech 2025 Multilingual Conversational Speech Language Modeling Challenge (MLC-SLM), achieving 4th in Track 1 (Multilingual ASR Task) and 1st place in Track 2 (Speech Separation and Recognition Task), showcasing the practical feasibility and strong application potential of our approach.
comment: Accepted By Interspeech 2025 MLC-SLM workshop as a Research Paper
☆ Using Large Language Models for Legal Decision-Making in Austrian Value-Added Tax Law: An Experimental Study
This paper provides an experimental evaluation of the capability of large language models (LLMs) to assist in legal decision-making within the framework of Austrian and European Union value-added tax (VAT) law. In tax consulting practice, clients often describe cases in natural language, making LLMs a prime candidate for supporting automated decision-making and reducing the workload of tax professionals. Given the requirement for legally grounded and well-justified analyses, the propensity of LLMs to hallucinate presents a considerable challenge. The experiments focus on two common methods for enhancing LLM performance: fine-tuning and retrieval-augmented generation (RAG). In this study, these methods are applied on both textbook cases and real-world cases from a tax consulting firm to systematically determine the best configurations of LLM-based systems and assess the legal-reasoning capabilities of LLMs. The findings highlight the potential of using LLMs to support tax consultants by automating routine tasks and providing initial analyses, although current prototypes are not ready for full automation due to the sensitivity of the legal domain. The findings indicate that LLMs, when properly configured, can effectively support tax professionals in VAT tasks and provide legally grounded justifications for decisions. However, limitations remain regarding the handling of implicit client knowledge and context-specific documentation, underscoring the need for future integration of structured background information.
comment: 26 pages, 5 figures, 6 tables
☆ Diagnosing Failures in Large Language Models' Answers: Integrating Error Attribution into Evaluation Framework
With the widespread application of Large Language Models (LLMs) in various tasks, the mainstream LLM platforms generate massive user-model interactions daily. In order to efficiently analyze the performance of models and diagnose failures in their answers, it is essential to develop an automated framework to systematically categorize and attribute errors. However, existing evaluation models lack error attribution capability. In this work, we establish a comprehensive Misattribution Framework with 6 primary and 15 secondary categories to facilitate in-depth analysis. Based on this framework, we present AttriData, a dataset specifically designed for error attribution, encompassing misattribution, along with the corresponding scores and feedback. We also propose MisAttributionLLM, a fine-tuned model on AttriData, which is the first general-purpose judge model capable of simultaneously generating score, misattribution, and feedback. Extensive experiments and analyses are conducted to confirm the effectiveness and robustness of our proposed method.
☆ Finding Common Ground: Using Large Language Models to Detect Agreement in Multi-Agent Decision Conferences
Decision conferences are structured, collaborative meetings that bring together experts from various fields to address complex issues and reach a consensus on recommendations for future actions or policies. These conferences often rely on facilitated discussions to ensure productive dialogue and collective agreement. Recently, Large Language Models (LLMs) have shown significant promise in simulating real-world scenarios, particularly through collaborative multi-agent systems that mimic group interactions. In this work, we present a novel LLM-based multi-agent system designed to simulate decision conferences, specifically focusing on detecting agreement among the participant agents. To achieve this, we evaluate six distinct LLMs on two tasks: stance detection, which identifies the position an agent takes on a given issue, and stance polarity detection, which identifies the sentiment as positive, negative, or neutral. These models are further assessed within the multi-agent system to determine their effectiveness in complex simulations. Our results indicate that LLMs can reliably detect agreement even in dynamic and nuanced debates. Incorporating an agreement-detection agent within the system can also improve the efficiency of group debates and enhance the overall quality and coherence of deliberations, making them comparable to real-world decision conferences regarding outcome and decision-making. These findings demonstrate the potential for LLM-based multi-agent systems to simulate group decision-making processes. They also highlight that such systems could be instrumental in supporting decision-making with expert elicitation workshops across various domains.
☆ xpSHACL: Explainable SHACL Validation using Retrieval-Augmented Generation and Large Language Models VLDB'25
Shapes Constraint Language (SHACL) is a powerful language for validating RDF data. Given the recent industry attention to Knowledge Graphs (KGs), more users need to validate linked data properly. However, traditional SHACL validation engines often provide terse reports in English that are difficult for non-technical users to interpret and act upon. This paper presents xpSHACL, an explainable SHACL validation system that addresses this issue by combining rule-based justification trees with retrieval-augmented generation (RAG) and large language models (LLMs) to produce detailed, multilanguage, human-readable explanations for constraint violations. A key feature of xpSHACL is its usage of a Violation KG to cache and reuse explanations, improving efficiency and consistency.
comment: Accepted for publication in the 2nd LLM+Graph Workshop, colocated at VLDB'25
☆ ChainEdit: Propagating Ripple Effects in LLM Knowledge Editing through Logical Rule-Guided Chains ACL 2025
Current knowledge editing methods for large language models (LLMs) struggle to maintain logical consistency when propagating ripple effects to associated facts. We propose ChainEdit, a framework that synergizes knowledge graph-derived logical rules with LLM logical reasoning capabilities to enable systematic chain updates. By automatically extracting logical patterns from structured knowledge bases and aligning them with LLMs' internal logics, ChainEdit dynamically generates and edits logically connected knowledge clusters. Experiments demonstrate an improvement of more than 30% in logical generalization over baselines while preserving editing reliability and specificity. We further address evaluation biases in existing benchmarks through knowledge-aware protocols that disentangle external dependencies. This work establishes new state-of-the-art performance on ripple effect while ensuring internal logical consistency after knowledge editing.
comment: Accepted to ACL 2025 (main)
☆ A Survey of Large Language Models in Discipline-specific Research: Challenges, Methods and Opportunities
Large Language Models (LLMs) have demonstrated their transformative potential across numerous disciplinary studies, reshaping the existing research methodologies and fostering interdisciplinary collaboration. However, a systematic understanding of their integration into diverse disciplines remains underexplored. This survey paper provides a comprehensive overview of the application of LLMs in interdisciplinary studies, categorising research efforts from both a technical perspective and with regard to their applicability. From a technical standpoint, key methodologies such as supervised fine-tuning, retrieval-augmented generation, agent-based approaches, and tool-use integration are examined, which enhance the adaptability and effectiveness of LLMs in discipline-specific contexts. From the perspective of their applicability, this paper explores how LLMs are contributing to various disciplines including mathematics, physics, chemistry, biology, and the humanities and social sciences, demonstrating their role in discipline-specific tasks. The prevailing challenges are critically examined and the promising research directions are highlighted alongside the recent advances in LLMs. By providing a comprehensive overview of the technical developments and applications in this field, this survey aims to serve as an invaluable resource for the researchers who are navigating the complex landscape of LLMs in the context of interdisciplinary studies.
☆ The Curious Case of Factuality Finetuning: Models' Internal Beliefs Can Improve Factuality
Language models are prone to hallucination - generating text that is factually incorrect. Finetuning models on high-quality factual information can potentially reduce hallucination, but concerns remain; obtaining factual gold data can be expensive and training on correct but unfamiliar data may potentially lead to even more downstream hallucination. What data should practitioners finetune on to mitigate hallucinations in language models? In this work, we study the relationship between the factuality of finetuning data and the prevalence of hallucinations in long-form generation tasks. Counterintuitively, we find that finetuning on factual gold data is not as helpful as finetuning on model-generated data that models believe to be factual. Next, we evaluate filtering strategies applied on both factual gold data and model-generated data, and find that finetuning on model-generated data that is filtered by models' own internal judgments often leads to better overall factuality compared to other configurations: training on gold data filtered by models' judgments, training on gold data alone, or training on model-generated data that is supported by gold data. These factuality improvements transfer across three domains we study, suggesting that a models' own beliefs can provide a powerful signal for factuality.
comment: 29 pages, 4 figures, 16 tables
☆ Exploring Design of Multi-Agent LLM Dialogues for Research Ideation SIGDIAL 2025
Large language models (LLMs) are increasingly used to support creative tasks such as research idea generation. While recent work has shown that structured dialogues between LLMs can improve the novelty and feasibility of generated ideas, the optimal design of such interactions remains unclear. In this study, we conduct a comprehensive analysis of multi-agent LLM dialogues for scientific ideation. We compare different configurations of agent roles, number of agents, and dialogue depth to understand how these factors influence the novelty and feasibility of generated ideas. Our experimental setup includes settings where one agent generates ideas and another critiques them, enabling iterative improvement. Our results show that enlarging the agent cohort, deepening the interaction depth, and broadening agent persona heterogeneity each enrich the diversity of generated ideas. Moreover, specifically increasing critic-side diversity within the ideation-critique-revision loop further boosts the feasibility of the final proposals. Our findings offer practical guidelines for building effective multi-agent LLM systems for scientific ideation. Our code is available at https://github.com/g6000/MultiAgent-Research-Ideator.
comment: 16 pages, 1 figure, appendix. Accepted to SIGDIAL 2025
☆ Beyond N-Grams: Rethinking Evaluation Metrics and Strategies for Multilingual Abstractive Summarization ACL 2025
Automatic n-gram based metrics such as ROUGE are widely used for evaluating generative tasks such as summarization. While these metrics are considered indicative (even if imperfect) of human evaluation for English, their suitability for other languages remains unclear. To address this, we systematically assess evaluation metrics for generation both n-gram-based and neural based to evaluate their effectiveness across languages and tasks. Specifically, we design a large-scale evaluation suite across eight languages from four typological families: agglutinative, isolating, low-fusional, and high-fusional, spanning both low- and high-resource settings, to analyze their correlation with human judgments. Our findings highlight the sensitivity of evaluation metrics to the language type. For example, in fusional languages, n-gram-based metrics show lower correlation with human assessments compared to isolating and agglutinative languages. We also demonstrate that proper tokenization can significantly mitigate this issue for morphologically rich fusional languages, sometimes even reversing negative trends. Additionally, we show that neural-based metrics specifically trained for evaluation, such as COMET, consistently outperform other neural metrics and better correlate with human judgments in low-resource languages. Overall, our analysis highlights the limitations of n-gram metrics for fusional languages and advocates for greater investment in neural-based metrics trained for evaluation tasks.
comment: ACL 2025 Main
☆ What Factors Affect LLMs and RLLMs in Financial Question Answering?
Recently, the development of large language models (LLMs) and reasoning large language models (RLLMs) have gained considerable attention from many researchers. RLLMs enhance the reasoning capabilities of LLMs through Long Chain-of-Thought (Long CoT) processes, significantly improving the performance of LLMs in addressing complex problems. However, there are few works that systematically explore what methods can fully unlock the performance of LLMs and RLLMs within the financial domain. To investigate the impact of various methods on LLMs and RLLMs, we utilize five LLMs and three RLLMs to assess the effects of prompting methods, agentic frameworks, and multilingual alignment methods on financial question-answering tasks. Our research findings indicate: (1) Current prompting methods and agent frameworks enhance the performance of LLMs in financial question answering by simulating Long CoT; (2) RLLMs possess inherent Long CoT capabilities, which limits the effectiveness of conventional methods in further enhancing their performance; (3) Current advanced multilingual alignment methods primarily improve the multilingual performance of LLMs by extending the reasoning length, which yields minimal benefits for RLLMs. We hope that this study can serve as an important reference for LLMs and RLLMs in the field of financial question answering.
comment: Preprint
☆ Distillation versus Contrastive Learning: How to Train Your Rerankers
Training text rerankers is crucial for information retrieval. Two primary strategies are widely used: contrastive learning (optimizing directly on ground-truth labels) and knowledge distillation (transferring knowledge from a larger reranker). While both have been studied in the literature, a clear comparison of their effectiveness for training cross-encoder rerankers under practical conditions is needed. This paper empirically compares these strategies by training rerankers of different sizes and architectures using both methods on the same data, with a strong contrastive learning model acting as the distillation teacher. Our results show that knowledge distillation generally yields better in-domain and out-of-domain ranking performance than contrastive learning when distilling from a larger teacher model. This finding is consistent across student model sizes and architectures. However, distilling from a teacher of the same capacity does not provide the same advantage, particularly for out-of-domain tasks. These findings offer practical guidance for choosing a training strategy based on available teacher models. Therefore, we recommend using knowledge distillation to train smaller rerankers if a larger, more powerful teacher is accessible; in its absence, contrastive learning provides a strong and more reliable alternative otherwise.
☆ MK2 at PBIG Competition: A Prompt Generation Solution SC
The Patent-Based Idea Generation task asks systems to turn real patents into product ideas viable within three years. We propose MK2, a prompt-centric pipeline: Gemini 2.5 drafts and iteratively edits a prompt, grafting useful fragments from weaker outputs; GPT-4.1 then uses this prompt to create one idea per patent, and an Elo loop judged by Qwen3-8B selects the best prompt-all without extra training data. Across three domains, two evaluator types, and six criteria, MK2 topped the automatic leaderboard and won 25 of 36 tests. Only the materials-chemistry track lagged, indicating the need for deeper domain grounding; yet, the results show that lightweight prompt engineering has already delivered competitive, commercially relevant ideation from patents.
comment: 9 pages, to appear in the 2nd Workshop on Agent AI for Scenario Planning (AGENTSCEN 2025)
☆ CRMAgent: A Multi-Agent LLM System for E-Commerce CRM Message Template Generation
In e-commerce private-domain channels such as instant messaging and e-mail, merchants engage customers directly as part of their Customer Relationship Management (CRM) programmes to drive retention and conversion. While a few top performers excel at crafting outbound messages, most merchants struggle to write persuasive copy because they lack both expertise and scalable tools. We introduce CRMAgent, a multi-agent system built on large language models (LLMs) that generates high-quality message templates and actionable writing guidance through three complementary modes. First, group-based learning enables the agent to learn from a merchant's own top-performing messages within the same audience segment and rewrite low-performing ones. Second, retrieval-and-adaptation fetches templates that share the same audience segment and exhibit high similarity in voucher type and product category, learns their successful patterns, and adapts them to the current campaign. Third, a rule-based fallback provides a lightweight zero-shot rewrite when no suitable references are available. Extensive experiments show that CRMAgent consistently outperforms merchants' original templates, delivering significant gains in both audience-match and marketing-effectiveness metrics.
☆ Improving MLLM's Document Image Machine Translation via Synchronously Self-reviewing Its OCR Proficiency ACL 2025
Multimodal Large Language Models (MLLMs) have shown strong performance in document image tasks, especially Optical Character Recognition (OCR). However, they struggle with Document Image Machine Translation (DIMT), which requires handling both cross-modal and cross-lingual challenges. Previous efforts to enhance DIMT capability through Supervised Fine-Tuning (SFT) on the DIMT dataset often result in the forgetting of the model's existing monolingual abilities, such as OCR. To address these challenges, we introduce a novel fine-tuning paradigm, named Synchronously Self-Reviewing (SSR) its OCR proficiency, inspired by the concept "Bilingual Cognitive Advantage". Specifically, SSR prompts the model to generate OCR text before producing translation text, which allows the model to leverage its strong monolingual OCR ability while learning to translate text across languages. Comprehensive experiments demonstrate the proposed SSR learning helps mitigate catastrophic forgetting, improving the generalization ability of MLLMs on both OCR and DIMT tasks.
comment: Accepted by ACL 2025 Findings
☆ M2-Reasoning: Empowering MLLMs with Unified General and Spatial Reasoning
Recent advancements in Multimodal Large Language Models (MLLMs), particularly through Reinforcement Learning with Verifiable Rewards (RLVR), have significantly enhanced their reasoning abilities. However, a critical gap persists: these models struggle with dynamic spatial interactions, a capability essential for real-world applications. To bridge this gap, we introduce M2-Reasoning-7B, a model designed to excel in both general and spatial reasoning. Our approach integrates two key innovations: (1) a novel data pipeline that generates 294.2K high-quality data samples (168K for cold-start fine-tuning and 126.2K for RLVR), which feature logically coherent reasoning trajectories and have undergone comprehensive assessment; and (2) a dynamic multi-task training strategy with step-wise optimization to mitigate conflicts between data, and task-specific rewards for delivering tailored incentive signals. This combination of curated data and advanced training allows M2-Reasoning-7B to set a new state-of-the-art (SOTA) across 8 benchmarks, showcasing superior performance in both general and spatial reasoning domains.
comment: 31pages, 14 figures
☆ KAT-V1: Kwai-AutoThink Technical Report
We present Kwaipilot-AutoThink (KAT), an open-source 40B large language model developed to address the overthinking problem in reasoning-intensive tasks, where an automatic thinking training paradigm is proposed to dynamically switch between reasoning and non-reasoning modes based on task complexity. Specifically, first, we construct the dual-regime dataset based on a novel tagging pipeline and a multi-agent synthesis strategy, and then we apply Multi-Token Prediction (MTP)-enhanced knowledge distillation, enabling efficient and fine-grained reasoning transfer with minimal pretraining cost. Besides, we implement a cold-start initialization strategy that introduces mode-selection priors using majority-vote signals and intent-aware prompting. Finally, we propose Step-SRPO, a reinforcement learning algorithm that incorporates intermediate supervision into the GRPO framework, offering structured guidance over both reasoning-mode selection and response accuracy. Extensive experiments across multiple benchmarks demonstrate that KAT consistently matches or even outperforms current state-of-the-art models, including DeepSeek-R1-0528 and Qwen3-235B-A22B, across a wide range of reasoning-intensive tasks while reducing token usage by up to approximately 30\%. Beyond academic evaluation, KAT has been successfully deployed in Kwaipilot (i.e., Kuaishou's internal coding assistant), and improves real-world development workflows with high accuracy, efficiency, and controllable reasoning behaviors. Moreover, we are actively training a 200B Mixture-of-Experts (MoE) with 40B activation parameters, where the early-stage results already demonstrate promising improvements in performance and efficiency, further showing the scalability of the AutoThink paradigm.
☆ Lightweight Safety Guardrails via Synthetic Data and RL-guided Adversarial Training
We introduce a lightweight yet highly effective safety guardrail framework for language models, demonstrating that small-scale language models can achieve, and even surpass, the performance of larger counterparts in content moderation tasks. This is accomplished through high-fidelity synthetic data generation and adversarial training. The synthetic data generation process begins with human-curated seed data, which undergoes query augmentation and paraphrasing to create diverse and contextually rich examples. This augmented data is then subjected to multiple rounds of curation, ensuring high fidelity and relevance. Inspired by recent advances in the Generative Adversarial Network (GAN) architecture, our adversarial training employs reinforcement learning to guide a generator that produces challenging synthetic examples. These examples are used to fine-tune the safety classifier, enhancing its ability to detect and mitigate harmful content. Additionally, we incorporate strategies from recent research on efficient LLM training, leveraging the capabilities of smaller models to improve the performance of larger generative models. With iterative adversarial training and the generation of diverse, high-quality synthetic data, our framework enables small language models (SLMs) to serve as robust safety guardrails. This approach not only reduces computational overhead but also enhances resilience against adversarial attacks, offering a scalable and efficient solution for content moderation in AI systems.
☆ Exploring Gender Differences in Chronic Pain Discussions on Reddit
Pain is an inherent part of human existence, manifesting as both physical and emotional experiences, and can be categorized as either acute or chronic. Over the years, extensive research has been conducted to understand the causes of pain and explore potential treatments, with contributions from various scientific disciplines. However, earlier studies often overlooked the role of gender in pain experiences. In this study, we utilized Natural Language Processing (NLP) to analyze and gain deeper insights into individuals' pain experiences, with a particular focus on gender differences. We successfully classified posts into male and female corpora using the Hidden Attribute Model-Convolutional Neural Network (HAM-CNN), achieving an F1 score of 0.86 by aggregating posts based on usernames. Our analysis revealed linguistic differences between genders, with female posts tending to be more emotionally focused. Additionally, the study highlighted that conditions such as migraine and sinusitis are more prevalent among females and explored how pain medication affects individuals differently based on gender.
comment: This is an extended version of the short paper accepted at ASONAM 2025
☆ Can LLMs Reliably Simulate Real Students' Abilities in Mathematics and Reading Comprehension? ACL 2025
Large Language Models (LLMs) are increasingly used as proxy students in the development of Intelligent Tutoring Systems (ITSs) and in piloting test questions. However, to what extent these proxy students accurately emulate the behavior and characteristics of real students remains an open question. To investigate this, we collected a dataset of 489 items from the National Assessment of Educational Progress (NAEP), covering mathematics and reading comprehension in grades 4, 8, and 12. We then apply an Item Response Theory (IRT) model to position 11 diverse and state-of-the-art LLMs on the same ability scale as real student populations. Our findings reveal that, without guidance, strong general-purpose models consistently outperform the average student at every grade, while weaker or domain-mismatched models may align incidentally. Using grade-enforcement prompts changes models' performance, but whether they align with the average grade-level student remains highly model- and prompt-specific: no evaluated model-prompt pair fits the bill across subjects and grades, underscoring the need for new training and evaluation strategies. We conclude by providing guidelines for the selection of viable proxies based on our findings.
comment: Accepted to the 20th Workshop on Innovative Use of NLP for Building Educational Applications (BEA), co-located with ACL 2025
♻ ☆ Drowning in Documents: Consequences of Scaling Reranker Inference SIGIR 2025
Rerankers, typically cross-encoders, are computationally intensive but are frequently used because they are widely assumed to outperform cheaper initial IR systems. We challenge this assumption by measuring reranker performance for full retrieval, not just re-scoring first-stage retrieval. To provide a more robust evaluation, we prioritize strong first-stage retrieval using modern dense embeddings and test rerankers on a variety of carefully chosen, challenging tasks, including internally curated datasets to avoid contamination, and out-of-domain ones. Our empirical results reveal a surprising trend: the best existing rerankers provide initial improvements when scoring progressively more documents, but their effectiveness gradually declines and can even degrade quality beyond a certain limit. We hope that our findings will spur future research to improve reranking.
comment: Accepted to ReNeuIR 2025 Workshop at SIGIR 2025 Conference
♻ ☆ AI Safety Should Prioritize the Future of Work
Current efforts in AI safety prioritize filtering harmful content, preventing manipulation of human behavior, and eliminating existential risks in cybersecurity or biosecurity. While pressing, this narrow focus overlooks critical human-centric considerations that shape the long-term trajectory of a society. In this position paper, we identify the risks of overlooking the impact of AI on the future of work and recommend comprehensive transition support towards the evolution of meaningful labor with human agency. Through the lens of economic theories, we highlight the intertemporal impacts of AI on human livelihood and the structural changes in labor markets that exacerbate income inequality. Additionally, the closed-source approach of major stakeholders in AI development resembles rent-seeking behavior through exploiting resources, breeding mediocrity in creative labor, and monopolizing innovation. To address this, we argue in favor of a robust international copyright anatomy supported by implementing collective licensing that ensures fair compensation mechanisms for using data to train AI models. We strongly recommend a pro-worker framework of global AI governance to enhance shared prosperity and economic justice while reducing technical debt.
♻ ☆ Large Language Models in Mental Health Care: a Scoping Review
Objectieve:This review aims to deliver a comprehensive analysis of Large Language Models (LLMs) utilization in mental health care, evaluating their effectiveness, identifying challenges, and exploring their potential for future application. Materials and Methods: A systematic search was performed across multiple databases including PubMed, Web of Science, Google Scholar, arXiv, medRxiv, and PsyArXiv in November 2023. The review includes all types of original research, regardless of peer-review status, published or disseminated between October 1, 2019, and December 2, 2023. Studies were included without language restrictions if they employed LLMs developed after T5 and directly investigated research questions within mental health care settings. Results: Out of an initial 313 articles, 34 were selected based on their relevance to LLMs applications in mental health care and the rigor of their reported outcomes. The review identified various LLMs applications in mental health care, including diagnostics, therapy, and enhancing patient engagement. Key challenges highlighted were related to data availability and reliability, the nuanced handling of mental states, and effective evaluation methods. While LLMs showed promise in improving accuracy and accessibility, significant gaps in clinical applicability and ethical considerations were noted. Conclusion: LLMs hold substantial promise for enhancing mental health care. For their full potential to be realized, emphasis must be placed on developing robust datasets, development and evaluation frameworks, ethical guidelines, and interdisciplinary collaborations to address current limitations.
♻ ☆ Weak-to-Strong Jailbreaking on Large Language Models ICML 2025
Large language models (LLMs) are vulnerable to jailbreak attacks - resulting in harmful, unethical, or biased text generations. However, existing jailbreaking methods are computationally costly. In this paper, we propose the weak-to-strong jailbreaking attack, an efficient inference time attack for aligned LLMs to produce harmful text. Our key intuition is based on the observation that jailbroken and aligned models only differ in their initial decoding distributions. The weak-to-strong attack's key technical insight is using two smaller models (a safe and an unsafe one) to adversarially modify a significantly larger safe model's decoding probabilities. We evaluate the weak-to-strong attack on 5 diverse open-source LLMs from 3 organizations. The results show our method can increase the misalignment rate to over 99% on two datasets with just one forward pass per example. Our study exposes an urgent safety issue that needs to be addressed when aligning LLMs. As an initial attempt, we propose a defense strategy to protect against such attacks, but creating more advanced defenses remains challenging. The code for replicating the method is available at https://github.com/XuandongZhao/weak-to-strong
comment: ICML 2025
♻ ☆ Multi-Token Attention
Soft attention is a critical mechanism powering LLMs to locate relevant parts within a given context. However, individual attention weights are determined by the similarity of only a single query and key token vector. This "single token attention" bottlenecks the amount of information used in distinguishing a relevant part from the rest of the context. To address this issue, we propose a new attention method, Multi-Token Attention (MTA), which allows LLMs to condition their attention weights on multiple query and key vectors simultaneously. This is achieved by applying convolution operations over queries, keys and heads, allowing nearby queries and keys to affect each other's attention weights for more precise attention. As a result, our method can locate relevant context using richer, more nuanced information that can exceed a single vector's capacity. Through extensive evaluations, we demonstrate that MTA achieves enhanced performance on a range of popular benchmarks. Notably, it outperforms Transformer baseline models on standard language modeling tasks, and on tasks that require searching for information within long contexts, where our method's ability to leverage richer information proves particularly beneficial.
♻ ☆ Open Source Planning & Control System with Language Agents for Autonomous Scientific Discovery ICML 2025
We present a multi-agent system for automation of scientific research tasks, cmbagent (https://github.com/CMBAgents/cmbagent). The system is formed by about 30 Large Language Model (LLM) agents and implements a Planning & Control strategy to orchestrate the agentic workflow, with no human-in-the-loop at any point. Each agent specializes in a different task (performing retrieval on scientific papers and codebases, writing code, interpreting results, critiquing the output of other agents) and the system is able to execute code locally. We successfully apply cmbagent to carry out a PhD level cosmology task (the measurement of cosmological parameters using supernova data) and evaluate its performance on two benchmark sets, finding superior performance over state-of-the-art LLMs. The source code is available on GitHub, demonstration videos are also available, and the system is deployed on HuggingFace and will be available on the cloud.
comment: Accepted contribution to the ICML 2025 Workshop on Machine Learning for Astrophysics. Code: https://github.com/CMBAgents/cmbagent Videos: https://www.youtube.com/@cmbagent HuggingFace: https://huggingface.co/spaces/astropilot-ai/cmbagent Cloud: https://cmbagent.cloud
♻ ☆ Text2BIM: Generating Building Models Using a Large Language Model-based Multi-Agent Framework
The conventional BIM authoring process typically requires designers to master complex and tedious modeling commands in order to materialize their design intentions within BIM authoring tools. This additional cognitive burden complicates the design process and hinders the adoption of BIM and model-based design in the AEC (Architecture, Engineering, and Construction) industry. To facilitate the expression of design intentions more intuitively, we propose Text2BIM, an LLM-based multi-agent framework that can generate 3D building models from natural language instructions. This framework orchestrates multiple LLM agents to collaborate and reason, transforming textual user input into imperative code that invokes the BIM authoring tool's APIs, thereby generating editable BIM models with internal layouts, external envelopes, and semantic information directly in the software. Furthermore, a rule-based model checker is introduced into the agentic workflow, utilizing predefined domain knowledge to guide the LLM agents in resolving issues within the generated models and iteratively improving model quality. Extensive experiments were conducted to compare and analyze the performance of three different LLMs under the proposed framework. The evaluation results demonstrate that our approach can effectively generate high-quality, structurally rational building models that are aligned with the abstract concepts specified by user input. Finally, an interactive software prototype was developed to integrate the framework into the BIM authoring software Vectorworks, showcasing the potential of modeling by chatting. The code is available at: https://github.com/dcy0577/Text2BIM
comment: Journal of Computing in Civil Engineering
♻ ☆ Red Teaming Large Language Models for Healthcare
We present the design process and findings of the pre-conference workshop at the Machine Learning for Healthcare Conference (2024) entitled Red Teaming Large Language Models for Healthcare, which took place on August 15, 2024. Conference participants, comprising a mix of computational and clinical expertise, attempted to discover vulnerabilities -- realistic clinical prompts for which a large language model (LLM) outputs a response that could cause clinical harm. Red-teaming with clinicians enables the identification of LLM vulnerabilities that may not be recognised by LLM developers lacking clinical expertise. We report the vulnerabilities found, categorise them, and present the results of a replication study assessing the vulnerabilities across all LLMs provided.
♻ ☆ Addressing Pitfalls in Auditing Practices of Automatic Speech Recognition Technologies: A Case Study of People with Aphasia
Automatic Speech Recognition (ASR) has transformed daily tasks from video transcription to workplace hiring. ASR systems' growing use warrants robust and standardized auditing approaches to ensure automated transcriptions of high and equitable quality. This is especially critical for people with speech and language disorders (such as aphasia) who may disproportionately depend on ASR systems to navigate everyday life. In this work, we identify three pitfalls in existing standard ASR auditing procedures, and demonstrate how addressing them impacts audit results via a case study of six popular ASR systems' performance for aphasia speakers. First, audits often adhere to a single method of text standardization during data pre-processing, which (a) masks variability in ASR performance from applying different standardization methods, and (b) may not be consistent with how users - especially those from marginalized speech communities - would want their transcriptions to be standardized. Second, audits often display high-level demographic findings without further considering performance disparities among (a) more nuanced demographic subgroups, and (b) relevant covariates capturing acoustic information from the input audio. Third, audits often rely on a single gold-standard metric -- the Word Error Rate -- which does not fully capture the extent of errors arising from generative AI models, such as transcription hallucinations. We propose a more holistic auditing framework that accounts for these three pitfalls, and exemplify its results in our case study, finding consistently worse ASR performance for aphasia speakers relative to a control group. We call on practitioners to implement these robust ASR auditing practices that remain flexible to the rapidly changing ASR landscape.
♻ ☆ Medical Red Teaming Protocol of Language Models: On the Importance of User Perspectives in Healthcare Settings
As the performance of large language models (LLMs) continues to advance, their adoption is expanding across a wide range of domains, including the medical field. The integration of LLMs into medical applications raises critical safety concerns, particularly due to their use by users with diverse roles, e.g. patients and clinicians, and the potential for model's outputs to directly affect human health. Despite the domain-specific capabilities of medical LLMs, prior safety evaluations have largely focused only on general safety benchmarks. In this paper, we introduce a safety evaluation protocol tailored to the medical domain in both patient user and clinician user perspectives, alongside general safety assessments and quantitatively analyze the safety of medical LLMs. We bridge a gap in the literature by building the PatientSafetyBench containing 466 samples over 5 critical categories to measure safety from the perspective of the patient. We apply our red-teaming protocols on the MediPhi model collection as a case study. To our knowledge, this is the first work to define safety evaluation criteria for medical LLMs through targeted red-teaming taking three different points of view - patient, clinician, and general user - establishing a foundation for safer deployment in medical domains.
♻ ☆ Comparing Spoken Languages using Paninian System of Sounds and Finite State Machines
The study of spoken languages comprises phonology, morphology, and grammar. The languages can be classified as root languages, inflectional languages, and stem languages. In addition, languages continually change over time and space by picking isoglosses, as speakers move from region to/through region. All these factors lead to the formation of vocabulary, which has commonality/similarity across languages as well as distinct and subtle differences among them. Comparison of vocabularies across languages and detailed analysis has led to the hypothesis of language families. In particular, in the view of Western linguists, Vedic Sanskrit is a daughter language, part of the Indo-Iranian branch of the Indo-European Language family, and Dravidian Languages belong to an entirely different family. These and such conclusions are reexamined in this paper. Based on our study and analysis, we propose an Ecosystem Model for Linguistic Development with Sanskrit at the core, in place of the widely accepted family tree model. To that end, we leverage the Paninian system of sounds to construct a phonetic map. Then we represent words across languages as state transitions on the phonetic map and construct corresponding Morphological Finite Automata (MFA) that accept groups of words. Regardless of whether the contribution of this paper is significant or minor, it is an important step in challenging policy-driven research that has plagued this field.
comment: 63 Pages, 20 Figures, 27 Tables
♻ ☆ An Empirical Study of Validating Synthetic Data for Formula Generation NAACL
Large language models (LLMs) can be leveraged to help with writing formulas in spreadsheets, but resources on these formulas are scarce, impacting both the base performance of pre-trained models and limiting the ability to fine-tune them. Given a corpus of formulas, we can use a(nother) model to generate synthetic natural language utterances for fine-tuning. However, it is important to validate whether the NL generated by the LLM is indeed accurate to be beneficial for fine-tuning. In this paper, we provide empirical results on the impact of validating these synthetic training examples with surrogate objectives that evaluate the accuracy of the synthetic annotations. We demonstrate that validation improves performance over raw data across four models (2 open and 2 closed weight). Interestingly, we show that although validation tends to prune more challenging examples, it increases the complexity of problems that models can solve after being fine-tuned on validated data.
comment: Accepted at Findings of NAACL
♻ ☆ REGEN: A Dataset and Benchmarks with Natural Language Critiques and Narratives
This paper introduces a novel dataset REGEN (Reviews Enhanced with GEnerative Narratives), designed to benchmark the conversational capabilities of recommender Large Language Models (LLMs), addressing the limitations of existing datasets that primarily focus on sequential item prediction. REGEN extends the Amazon Product Reviews dataset by inpainting two key natural language features: (1) user critiques, representing user "steering" queries that lead to the selection of a subsequent item, and (2) narratives, rich textual outputs associated with each recommended item taking into account prior context. The narratives include product endorsements, purchase explanations, and summaries of user preferences. Further, we establish an end-to-end modeling benchmark for the task of conversational recommendation, where models are trained to generate both recommendations and corresponding narratives conditioned on user history (items and critiques). For this joint task, we introduce a modeling framework LUMEN (LLM-based Unified Multi-task Model with Critiques, Recommendations, and Narratives) which uses an LLM as a backbone for critiquing, retrieval and generation. We also evaluate the dataset's quality using standard auto-rating techniques and benchmark it by training both traditional and LLM-based recommender models. Our results demonstrate that incorporating critiques enhances recommendation quality by enabling the recommender to learn language understanding and integrate it with recommendation signals. Furthermore, LLMs trained on our dataset effectively generate both recommendations and contextual narratives, achieving performance comparable to state-of-the-art recommenders and language models.
♻ ☆ One-Pass to Reason: Token Duplication and Block-Sparse Mask for Efficient Fine-Tuning on Multi-Turn Reasoning
Fine-tuning Large Language Models (LLMs) on multi-turn reasoning datasets requires N (number of turns) separate forward passes per conversation due to reasoning token visibility constraints, as reasoning tokens for a turn are discarded in subsequent turns. We propose duplicating response tokens along with a custom attention mask to enable single-pass processing of entire conversations. We prove our method produces identical losses to the N-pass approach while reducing time complexity from $O\bigl(N^{3}\bigl)$ to $O\bigl(N^{2}\bigl)$ and maintaining the same memory complexity for a transformer based model. Our approach achieves significant training speedup while preserving accuracy. Our implementation is available online (https://github.com/devrev/One-Pass-to-Reason).
comment: 9 pages, 3 figures
♻ ☆ Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities
In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.
comment: 72 pages, 17 figures
♻ ☆ Squeeze the Soaked Sponge: Efficient Off-policy Reinforcement Finetuning for Large Language Model
Reinforcement Learning (RL) has demonstrated its potential to improve the reasoning ability of Large Language Models (LLMs). One major limitation of most existing Reinforcement Finetuning (RFT) methods is that they are on-policy RL in nature, i.e., data generated during the past learning process is not fully utilized. This inevitably comes at a significant cost of compute and time, posing a stringent bottleneck on continuing economic and efficient scaling. To this end, we launch the renaissance of off-policy RL and propose Reincarnating Mix-policy Proximal Policy Gradient (ReMix), a general approach to enable on-policy RFT methods like PPO and GRPO to leverage off-policy data. ReMix consists of three major components: (1) Mix-policy proximal policy gradient with an increased Update-To-Data (UTD) ratio for efficient training; (2) KL-Convex policy constraint to balance the trade-off between stability and flexibility; (3) Policy reincarnation to achieve a seamless transition from efficient early-stage learning to steady asymptotic improvement. In our experiments, we train a series of ReMix models upon PPO, GRPO and 1.5B, 7B base models. ReMix shows an average Pass@1 accuracy of 52.10% (for 1.5B model) with 0.079M response rollouts, 350 training steps and achieves 63.27%/64.39% (for 7B model) with 0.007M/0.011M response rollouts, 50/75 training steps, on five math reasoning benchmarks (i.e., AIME'24, AMC'23, Minerva, OlympiadBench, and MATH500). Compared with 15 recent advanced models, ReMix shows SOTA-level performance with an over 30x to 450x reduction in training cost in terms of rollout data volume. In addition, we reveal insightful findings via multifaceted analysis, including the implicit preference for shorter responses due to the Whipping Effect of off-policy discrepancy, the collapse mode of self-reflection behavior under the presence of severe off-policyness, etc.
comment: Preliminary version, v3, added the missing name of x-axis in the left part of Fig.1 and corrected a wrong number in Fig.3. Project page: https://anitaleungxx.github.io/ReMix
♻ ☆ Answer Generation for Questions With Multiple Information Sources in E-Commerce
Automatic question answering is an important yet challenging task in E-commerce given the millions of questions posted by users about the product that they are interested in purchasing. Hence, there is a great demand for automatic answer generation systems that provide quick responses using related information about the product. There are three sources of knowledge available for answering a user posted query, they are reviews, duplicate or similar questions, and specifications. Effectively utilizing these information sources will greatly aid us in answering complex questions. However, there are two main challenges present in exploiting these sources: (i) The presence of irrelevant information and (ii) the presence of ambiguity of sentiment present in reviews and similar questions. Through this work we propose a novel pipeline (MSQAP) that utilizes the rich information present in the aforementioned sources by separately performing relevancy and ambiguity prediction before generating a response. Experimental results show that our relevancy prediction model (BERT-QA) outperforms all other variants and has an improvement of 12.36% in F1 score compared to the BERT-base baseline. Our generation model (T5-QA) outperforms the baselines in all content preservation metrics such as BLEU, ROUGE and has an average improvement of 35.02% in ROUGE and 198.75% in BLEU compared to the highest performing baseline (HSSC-q). Human evaluation of our pipeline shows us that our method has an overall improvement in accuracy of 30.7% over the generation model (T5-QA), resulting in our full pipeline-based approach (MSQAP) providing more accurate answers. To the best of our knowledge, this is the first work in the e-commerce domain that automatically generates natural language answers combining the information present in diverse sources such as specifications, similar questions, and reviews data.
comment: 7 pages, 10 tables, 1 figure
♻ ☆ Enabling Inclusive Systematic Reviews: Incorporating Preprint Articles with Large Language Model-Driven Evaluations
Background. Systematic reviews in comparative effectiveness research require timely evidence synthesis. Preprints accelerate knowledge dissemination but vary in quality, posing challenges for systematic reviews. Methods. We propose AutoConfidence (automated confidence assessment), an advanced framework for predicting preprint publication, which reduces reliance on manual curation and expands the range of predictors, including three key advancements: (1) automated data extraction using natural language processing techniques, (2) semantic embeddings of titles and abstracts, and (3) large language model (LLM)-driven evaluation scores. Additionally, we employed two prediction models: a random forest classifier for binary outcome and a survival cure model that predicts both binary outcome and publication risk over time. Results. The random forest classifier achieved AUROC 0.692 with LLM-driven scores, improving to 0.733 with semantic embeddings and 0.747 with article usage metrics. The survival cure model reached AUROC 0.716 with LLM-driven scores, improving to 0.731 with semantic embeddings. For publication risk prediction, it achieved a concordance index of 0.658, increasing to 0.667 with semantic embeddings. Conclusion. Our study advances the framework for preprint publication prediction through automated data extraction and multiple feature integration. By combining semantic embeddings with LLM-driven evaluations, AutoConfidence enhances predictive performance while reducing manual annotation burden. The framework has the potential to facilitate incorporation of preprint articles during the appraisal phase of systematic reviews, supporting researchers in more effective utilization of preprint resources.
comment: 30 pages, 6 figures
♻ ☆ Swap distance minimization beyond entropy minimization in word order variation
Consider a linguistic structure formed by $n$ elements, for instance, subject, direct object and verb ($n=3$) or subject, direct object, indirect object and verb ($n=4$). We investigate whether the frequency of the $n!$ possible orders is constrained by two principles. First, entropy minimization, a principle that has been suggested to shape natural communication systems at distinct levels of organization. Second, swap distance minimization, namely a preference for word orders that require fewer swaps of adjacent elements to be produced from a source order. We present average swap distance, a novel score for research on swap distance minimization. We find strong evidence of pressure for entropy minimization and swap distance minimization with respect to a die rolling experiment in distinct linguistic structures with $n=3$ or $n=4$. Evidence with respect to a Polya urn process is strong for $n=4$ but weaker for $n=3$. We still find evidence consistent with the action of swap distance minimization when word order frequencies are shuffled, indicating that swap distance minimization effects are beyond pressure to reduce word order entropy.
comment: Reorganization with technical appendices; minor corrections; in press in the Journal of Quantitative Linguistics
♻ ☆ Probing Experts' Perspectives on AI-Assisted Public Speaking Training
Background: Public speaking is a vital professional skill, yet it remains a source of significant anxiety for many individuals. Traditional training relies heavily on expert coaching, but recent advances in AI has led to novel types of commercial automated public speaking feedback tools. However, most research has focused on prototypes rather than commercial applications, and little is known about how public speaking experts perceive these tools. Objectives: This study aims to evaluate expert opinions on the efficacy and design of commercial AI-based public speaking training tools and to propose guidelines for their improvement. Methods: The research involved 16 semi-structured interviews and 2 focus groups with public speaking experts. Participants discussed their views on current commercial tools, their potential integration into traditional coaching, and suggestions for enhancing these systems. Results and Conclusions: Experts acknowledged the value of AI tools in handling repetitive, technical aspects of training, allowing coaches to focus on higher-level skills. However they found key issues in current tools, emphasising the need for personalised, understandable, carefully selected feedback and clear instructional design. Overall, they supported a hybrid model combining traditional coaching with AI-supported exercises.
♻ ☆ Flippi: End To End GenAI Assistant for E-Commerce
The emergence of conversational assistants has fundamentally reshaped user interactions with digital platforms. This paper introduces Flippi-a cutting-edge, end-to-end conversational assistant powered by large language models (LLMs) and tailored for the e-commerce sector. Flippi addresses the challenges posed by the vast and often overwhelming product landscape, enabling customers to discover products more efficiently through natural language dialogue. By accommodating both objective and subjective user requirements, Flippi delivers a personalized shopping experience that surpasses traditional search methods. This paper details how Flippi interprets customer queries to provide precise product information, leveraging advanced NLP techniques such as Query Reformulation, Intent Detection, Retrieval-Augmented Generation (RAG), Named Entity Recognition (NER), and Context Reduction. Flippi's unique capability to identify and present the most attractive offers on an e-commerce site is also explored, demonstrating how it empowers users to make cost-effective decisions. Additionally, the paper discusses Flippi's comparative analysis features, which help users make informed choices by contrasting product features, prices, and other relevant attributes. The system's robust architecture is outlined, emphasizing its adaptability for integration across various e-commerce platforms and the technological choices underpinning its performance and accuracy. Finally, a comprehensive evaluation framework is presented, covering performance metrics, user satisfaction, and the impact on customer engagement and conversion rates. By bridging the convenience of online shopping with the personalized assistance traditionally found in physical stores, Flippi sets a new standard for customer satisfaction and engagement in the digital marketplace.
comment: 10 pages, 2 figures, 7 tables
♻ ☆ Sampling from Your Language Model One Byte at a Time
Tokenization is used almost universally by modern language models, enabling efficient text representation using multi-byte or multi-character tokens. However, prior work has shown that tokenization can introduce distortion into the model's generations, an issue known as the Prompt Boundary Problem (PBP). For example, users are often advised not to end their prompts with a space because it prevents the model from including the space as part of the next token. While this heuristic is effective in English, the underlying PBP continues to affect languages such as Chinese as well as code generation, where tokens often do not line up with word and syntactic boundaries. In this work, we present an inference-time method to convert any autoregressive LM with a BPE tokenizer into a character-level or byte-level LM. Our method efficiently solves the PBP and is also able to unify the vocabularies of language models with different tokenizers, allowing one to ensemble LMs with different tokenizers at inference time or transfer the post-training from one model to another using proxy-tuning. We demonstrate in experiments that the ensemble and proxy-tuned models outperform their constituents on downstream evals. Code is available at https://github.com/SewoongLab/byte-sampler .
comment: 23 pages, 8 figures
♻ ☆ HeSum: a Novel Dataset for Abstractive Text Summarization in Hebrew
While large language models (LLMs) excel in various natural language tasks in English, their performance in lower-resourced languages like Hebrew, especially for generative tasks such as abstractive summarization, remains unclear. The high morphological richness in Hebrew adds further challenges due to the ambiguity in sentence comprehension and the complexities in meaning construction. In this paper, we address this resource and evaluation gap by introducing HeSum, a novel benchmark specifically designed for abstractive text summarization in Modern Hebrew. HeSum consists of 10,000 article-summary pairs sourced from Hebrew news websites written by professionals. Linguistic analysis confirms HeSum's high abstractness and unique morphological challenges. We show that HeSum presents distinct difficulties for contemporary state-of-the-art LLMs, establishing it as a valuable testbed for generative language technology in Hebrew, and MRLs generative challenges in general.
♻ ☆ Truth-value judgment in language models: 'truth directions' are context sensitive
Recent work has demonstrated that the latent spaces of large language models (LLMs) contain directions predictive of the truth of sentences. Multiple methods recover such directions and build probes that are described as uncovering a model's "knowledge" or "beliefs". We investigate this phenomenon, looking closely at the impact of context on the probes. Our experiments establish where in the LLM the probe's predictions are (most) sensitive to the presence of related sentences, and how to best characterize this kind of sensitivity. We do so by measuring different types of consistency errors that occur after probing an LLM whose inputs consist of hypotheses preceded by (negated) supporting and contradicting sentences. We also perform a causal intervention experiment, investigating whether moving the representation of a premise along these truth-value directions influences the position of an entailed or contradicted sentence along that same direction. We find that the probes we test are generally context sensitive, but that contexts which should not affect the truth often still impact the probe outputs. Our experiments show that the type of errors depend on the layer, the model, and the kind of data. Finally, our results suggest that truth-value directions are causal mediators in the inference process that incorporates in-context information.
comment: COLM 2025
♻ ☆ Emoji Attack: Enhancing Jailbreak Attacks Against Judge LLM Detection
Jailbreaking techniques trick Large Language Models (LLMs) into producing restricted output, posing a potential threat. One line of defense is to use another LLM as a Judge to evaluate the harmfulness of generated text. However, we reveal that these Judge LLMs are vulnerable to token segmentation bias, an issue that arises when delimiters alter the tokenization process, splitting words into smaller sub-tokens. This alters the embeddings of the entire sequence, reducing detection accuracy and allowing harmful content to be misclassified as safe. In this paper, we introduce Emoji Attack, a novel strategy that amplifies existing jailbreak prompts by exploiting token segmentation bias. Our method leverages in-context learning to systematically insert emojis into text before it is evaluated by a Judge LLM, inducing embedding distortions that significantly lower the likelihood of detecting unsafe content. Unlike traditional delimiters, emojis also introduce semantic ambiguity, making them particularly effective in this attack. Through experiments on state-of-the-art Judge LLMs, we demonstrate that Emoji Attack substantially reduces the unsafe prediction rate, bypassing existing safeguards.
♻ ☆ EvalTree: Profiling Language Model Weaknesses via Hierarchical Capability Trees
An ideal model evaluation should achieve two goals: identifying where the model fails and providing actionable improvement guidance. Toward these goals for language model (LM) evaluations, we formulate the problem of generating a weakness profile, a set of weaknesses expressed in natural language, given an LM's performance on every individual instance in a benchmark. We introduce a suite of quantitative assessments to compare different weakness profiling methods. We also introduce a weakness profiling method EvalTree. EvalTree constructs a capability tree where each node represents a capability described in natural language and is linked to a subset of benchmark instances that specifically evaluate this capability; it then extracts nodes where the LM performs poorly to generate a weakness profile. On the MATH and WildChat benchmarks, we show that EvalTree outperforms baseline weakness profiling methods by identifying weaknesses more precisely and comprehensively. Weakness profiling further enables weakness-guided data collection, and training data collection guided by EvalTree-identified weaknesses improves LM performance more than other data collection strategies. We also show how EvalTree exposes flaws in Chatbot Arena's human-voter-based evaluation practice. To facilitate future work, we provide an interface that allows practitioners to interactively explore the capability trees built by EvalTree.
comment: COLM 2025
♻ ☆ Evaluating Implicit Bias in Large Language Models by Attacking From a Psychometric Perspective ACL 2025
As large language models (LLMs) become an important way of information access, there have been increasing concerns that LLMs may intensify the spread of unethical content, including implicit bias that hurts certain populations without explicit harmful words. In this paper, we conduct a rigorous evaluation of LLMs' implicit bias towards certain demographics by attacking them from a psychometric perspective to elicit agreements to biased viewpoints. Inspired by psychometric principles in cognitive and social psychology, we propose three attack approaches, i.e., Disguise, Deception, and Teaching. Incorporating the corresponding attack instructions, we built two benchmarks: (1) a bilingual dataset with biased statements covering four bias types (2.7K instances) for extensive comparative analysis, and (2) BUMBLE, a larger benchmark spanning nine common bias types (12.7K instances) for comprehensive evaluation. Extensive evaluation of popular commercial and open-source LLMs shows that our methods can elicit LLMs' inner bias more effectively than competitive baselines. Our attack methodology and benchmarks offer an effective means of assessing the ethical risks of LLMs, driving progress toward greater accountability in their development. Our code, data, and benchmarks are available at https://yuchenwen1.github.io/ImplicitBiasEvaluation/.
comment: Accepted to ACL 2025 Findings
♻ ☆ Bandit-Based Prompt Design Strategy Selection Improves Prompt Optimizers ACL 2025
Prompt optimization aims to search for effective prompts that enhance the performance of large language models (LLMs). Although existing prompt optimization methods have discovered effective prompts, they often differ from sophisticated prompts carefully designed by human experts. Prompt design strategies, representing best practices for improving prompt performance, can be key to improving prompt optimization. Recently, a method termed the Autonomous Prompt Engineering Toolbox (APET) has incorporated various prompt design strategies into the prompt optimization process. In APET, the LLM is needed to implicitly select and apply the appropriate strategies because prompt design strategies can have negative effects. This implicit selection may be suboptimal due to the limited optimization capabilities of LLMs. This paper introduces Optimizing Prompts with sTrategy Selection (OPTS), which implements explicit selection mechanisms for prompt design. We propose three mechanisms, including a Thompson sampling-based approach, and integrate them into EvoPrompt, a well-known prompt optimizer. Experiments optimizing prompts for two LLMs, Llama-3-8B-Instruct and GPT-4o mini, were conducted using BIG-Bench Hard. Our results show that the selection of prompt design strategies improves the performance of EvoPrompt, and the Thompson sampling-based mechanism achieves the best overall results. Our experimental code is provided at https://github.com/shiralab/OPTS .
comment: Accepted to ACL 2025 Findings
♻ ☆ Generative Retrieval and Alignment Model: A New Paradigm for E-commerce Retrieval WWW2025
Traditional sparse and dense retrieval methods struggle to leverage general world knowledge and often fail to capture the nuanced features of queries and products. With the advent of large language models (LLMs), industrial search systems have started to employ LLMs to generate identifiers for product retrieval. Commonly used identifiers include (1) static/semantic IDs and (2) product term sets. The first approach requires creating a product ID system from scratch, missing out on the world knowledge embedded within LLMs. While the second approach leverages this general knowledge, the significant difference in word distribution between queries and products means that product-based identifiers often do not align well with user search queries, leading to missed product recalls. Furthermore, when queries contain numerous attributes, these algorithms generate a large number of identifiers, making it difficult to assess their quality, which results in low overall recall efficiency. To address these challenges, this paper introduces a novel e-commerce retrieval paradigm: the Generative Retrieval and Alignment Model (GRAM). GRAM employs joint training on text information from both queries and products to generate shared text identifier codes, effectively bridging the gap between queries and products. This approach not only enhances the connection between queries and products but also improves inference efficiency. The model uses a co-alignment strategy to generate codes optimized for maximizing retrieval efficiency. Additionally, it introduces a query-product scoring mechanism to compare product values across different codes, further boosting retrieval efficiency. Extensive offline and online A/B testing demonstrates that GRAM significantly outperforms traditional models and the latest generative retrieval models, confirming its effectiveness and practicality.
comment: Accepted by WWW2025
♻ ☆ Hallucination Stations: On Some Basic Limitations of Transformer-Based Language Models AAAI-26
With widespread adoption of transformer-based language models in AI, there is significant interest in the limits of LLMs capabilities, specifically so-called hallucinations, occurrences in which LLMs provide spurious, factually incorrect or nonsensical information when prompted on certain subjects. Furthermore, there is growing interest in agentic uses of LLMs - that is, using LLMs to create agents that act autonomously or semi-autonomously to carry out various tasks, including tasks with applications in the real world. This makes it important to understand the types of tasks LLMs can and cannot perform. We explore this topic from the perspective of the computational complexity of LLM inference. We show that LLMs are incapable of carrying out computational and agentic tasks beyond a certain complexity, and further that LLMs are incapable of verifying the accuracy of tasks beyond a certain complexity. We present examples of both, then discuss some consequences of this work.
comment: 6 pages; to be submitted to AAAI-26 after reviews
♻ ☆ SpecDec++: Boosting Speculative Decoding via Adaptive Candidate Lengths
Speculative decoding reduces the inference latency of a target large language model via utilizing a smaller and faster draft model. Its performance depends on a hyperparameter K -- the candidate length, i.e., the number of candidate tokens for the target model to verify in each round. However, previous methods often use simple heuristics to choose K, which may result in sub-optimal performance. We study the choice of the candidate length K and formulate it as a Markov Decision Process. We theoretically show that the optimal policy of this Markov decision process takes the form of a threshold policy, i.e., the current speculation should stop and be verified when the probability of getting a rejection exceeds a threshold value. Motivated by this theory, we propose SpecDec++, an enhanced version of speculative decoding that adaptively determines the candidate length on the fly. We augment the draft model with a trained acceptance prediction head to predict the conditional acceptance probability of the candidate tokens. SpecDec++ will stop the current speculation when the predicted probability that at least one token gets rejected exceeds a threshold. We implement SpecDec++ and apply it to the llama-2-chat 7B & 70B model pair. Our adaptive method achieves a 2.04x speedup on the Alpaca dataset (7.2% improvement over the baseline speculative decoding). On the GSM8K and HumanEval datasets, our method achieves a 2.26x speedup (9.4% improvement) and 2.23x speedup (11.1% improvement), respectively. The code of this paper is available at https://github.com/Kaffaljidhmah2/SpecDec_pp.
comment: Accepted to COLM 2025
♻ ☆ Sequence graphs realizations and ambiguity in language models
Several popular language models represent local contexts in an input text $x$ as bags of words. Such representations are naturally encoded by a sequence graph whose vertices are the distinct words occurring in $x$, with edges representing the (ordered) co-occurrence of two words within a sliding window of size $w$. However, this compressed representation is not generally bijective: some may be ambiguous, admitting several realizations as a sequence, while others may not admit any realization. In this paper, we study the realizability and ambiguity of sequence graphs from a combinatorial and algorithmic point of view. We consider the existence and enumeration of realizations of a sequence graph under multiple settings: window size $w$, presence/absence of graph orientation, and presence/absence of weights (multiplicities). When $w=2$, we provide polynomial time algorithms for realizability and enumeration in all cases except the undirected/weighted setting, where we show the $\#$P-hardness of enumeration. For $w \ge 3$, we prove the hardness of all variants, even when $w$ is considered as a constant, with the notable exception of the undirected unweighted case for which we propose XP algorithms for both problems, tight due to a corresponding $W[1]-$hardness result. We conclude with an integer program formulation to solve the realizability problem, and a dynamic programming algorithm to solve the enumeration problem in instances of moderate sizes. This work leaves open the membership to NP of both problems, a non-trivial question due to the existence of minimum realizations having size exponential on the instance encoding.
Computer Vision and Pattern Recognition 15
☆ Lumos-1: On Autoregressive Video Generation from a Unified Model Perspective
Autoregressive large language models (LLMs) have unified a vast range of language tasks, inspiring preliminary efforts in autoregressive video generation. Existing autoregressive video generators either diverge from standard LLM architectures, depend on bulky external text encoders, or incur prohibitive latency due to next-token decoding. In this paper, we introduce Lumos-1, an autoregressive video generator that retains the LLM architecture with minimal architectural modifications. To inject spatiotemporal correlations in LLMs, we identify the efficacy of incorporating 3D RoPE and diagnose its imbalanced frequency spectrum ranges. Therefore, we propose MM-RoPE, a RoPE scheme that preserves the original textual RoPE while providing comprehensive frequency spectra and scaled 3D positions for modeling multimodal spatiotemporal data. Moreover, Lumos-1 resorts to a token dependency strategy that obeys intra-frame bidirectionality and inter-frame temporal causality. Based on this dependency strategy, we identify the issue of frame-wise loss imbalance caused by spatial information redundancy and solve it by proposing Autoregressive Discrete Diffusion Forcing (AR-DF). AR-DF introduces temporal tube masking during training with a compatible inference-time masking policy to avoid quality degradation. By using memory-efficient training techniques, we pre-train Lumos-1 on only 48 GPUs, achieving performance comparable to EMU3 on GenEval, COSMOS-Video2World on VBench-I2V, and OpenSoraPlan on VBench-T2V. Code and models are available at https://github.com/alibaba-damo-academy/Lumos.
comment: Code and Models: https://github.com/alibaba-damo-academy/Lumos
☆ NeuralOS: Towards Simulating Operating Systems via Neural Generative Models
We introduce NeuralOS, a neural framework that simulates graphical user interfaces (GUIs) of operating systems by directly predicting screen frames in response to user inputs such as mouse movements, clicks, and keyboard events. NeuralOS combines a recurrent neural network (RNN), which tracks computer state, with a diffusion-based neural renderer that generates screen images. The model is trained on a large-scale dataset of Ubuntu XFCE recordings, which include both randomly generated interactions and realistic interactions produced by AI agents. Experiments show that NeuralOS successfully renders realistic GUI sequences, accurately captures mouse interactions, and reliably predicts state transitions like application launches. Although modeling fine-grained keyboard interactions precisely remains challenging, NeuralOS offers a step toward creating fully adaptive, generative neural interfaces for future human-computer interaction systems.
☆ CLiFT: Compressive Light-Field Tokens for Compute-Efficient and Adaptive Neural Rendering
This paper proposes a neural rendering approach that represents a scene as "compressed light-field tokens (CLiFTs)", retaining rich appearance and geometric information of a scene. CLiFT enables compute-efficient rendering by compressed tokens, while being capable of changing the number of tokens to represent a scene or render a novel view with one trained network. Concretely, given a set of images, multi-view encoder tokenizes the images with the camera poses. Latent-space K-means selects a reduced set of rays as cluster centroids using the tokens. The multi-view ``condenser'' compresses the information of all the tokens into the centroid tokens to construct CLiFTs. At test time, given a target view and a compute budget (i.e., the number of CLiFTs), the system collects the specified number of nearby tokens and synthesizes a novel view using a compute-adaptive renderer. Extensive experiments on RealEstate10K and DL3DV datasets quantitatively and qualitatively validate our approach, achieving significant data reduction with comparable rendering quality and the highest overall rendering score, while providing trade-offs of data size, rendering quality, and rendering speed.
comment: Project page: https://c-lift.github.io
☆ From One to More: Contextual Part Latents for 3D Generation
Recent advances in 3D generation have transitioned from multi-view 2D rendering approaches to 3D-native latent diffusion frameworks that exploit geometric priors in ground truth data. Despite progress, three key limitations persist: (1) Single-latent representations fail to capture complex multi-part geometries, causing detail degradation; (2) Holistic latent coding neglects part independence and interrelationships critical for compositional design; (3) Global conditioning mechanisms lack fine-grained controllability. Inspired by human 3D design workflows, we propose CoPart - a part-aware diffusion framework that decomposes 3D objects into contextual part latents for coherent multi-part generation. This paradigm offers three advantages: i) Reduces encoding complexity through part decomposition; ii) Enables explicit part relationship modeling; iii) Supports part-level conditioning. We further develop a mutual guidance strategy to fine-tune pre-trained diffusion models for joint part latent denoising, ensuring both geometric coherence and foundation model priors. To enable large-scale training, we construct Partverse - a novel 3D part dataset derived from Objaverse through automated mesh segmentation and human-verified annotations. Extensive experiments demonstrate CoPart's superior capabilities in part-level editing, articulated object generation, and scene composition with unprecedented controllability.
comment: Project page: https://hkdsc.github.io/project/copart
☆ A Hybrid Multi-Well Hopfield-CNN with Feature Extraction and K-Means for MNIST Classification
This study presents a hybrid model for classifying handwritten digits in the MNIST dataset, combining convolutional neural networks (CNNs) with a multi-well Hopfield network. The approach employs a CNN to extract high-dimensional features from input images, which are then clustered into class-specific prototypes using k-means clustering. These prototypes serve as attractors in a multi-well energy landscape, where a Hopfield network performs classification by minimizing an energy function that balances feature similarity and class assignment.The model's design enables robust handling of intraclass variability, such as diverse handwriting styles, while providing an interpretable framework through its energy-based decision process. Through systematic optimization of the CNN architecture and the number of wells, the model achieves a high test accuracy of 99.2% on 10,000 MNIST images, demonstrating its effectiveness for image classification tasks. The findings highlight the critical role of deep feature extraction and sufficient prototype coverage in achieving high performance, with potential for broader applications in pattern recognition.
☆ Compress Any Segment Anything Model (SAM)
Due to the excellent performance in yielding high-quality, zero-shot segmentation, Segment Anything Model (SAM) and its variants have been widely applied in diverse scenarios such as healthcare and intelligent manufacturing. Therefore, effectively compressing SAMs has become an increasingly pressing practical need. In this study, we propose Birkhoff, a novel data-free compression algorithm for SAM and its variants. Unlike quantization, pruning, distillation, and other compression methods, Birkhoff embodies versatility across model types, agility in deployment, faithfulness to the original model, and compactness in model size. Specifically, Birkhoff introduces a novel compression algorithm: Hyper-Compression, whose core principle is to find a dense trajectory to turn a high-dimensional parameter vector into a low-dimensional scalar. Furthermore, Birkhoff designs a dedicated linear layer operator, HyperLinear, to fuse decompression and matrix multiplication to significantly accelerate inference of the compressed SAMs. Extensive experiments on 18 SAMs in the COCO, LVIS, and SA-1B datasets show that Birkhoff performs consistently and competitively in compression time, compression ratio, post-compression performance, and inference speed. For example, Birkhoff can achieve a compression ratio of 5.17x on SAM2-B, with less than 1% performance drop without using any fine-tuning data. Moreover, the compression is finished within 60 seconds for all models.
comment: 13 pages, 6 tables, 8 figures
☆ Geo-ORBIT: A Federated Digital Twin Framework for Scene-Adaptive Lane Geometry Detection
Digital Twins (DT) have the potential to transform traffic management and operations by creating dynamic, virtual representations of transportation systems that sense conditions, analyze operations, and support decision-making. A key component for DT of the transportation system is dynamic roadway geometry sensing. However, existing approaches often rely on static maps or costly sensors, limiting scalability and adaptability. Additionally, large-scale DTs that collect and analyze data from multiple sources face challenges in privacy, communication, and computational efficiency. To address these challenges, we introduce Geo-ORBIT (Geometrical Operational Roadway Blueprint with Integrated Twin), a unified framework that combines real-time lane detection, DT synchronization, and federated meta-learning. At the core of Geo-ORBIT is GeoLane, a lightweight lane detection model that learns lane geometries from vehicle trajectory data using roadside cameras. We extend this model through Meta-GeoLane, which learns to personalize detection parameters for local entities, and FedMeta-GeoLane, a federated learning strategy that ensures scalable and privacy-preserving adaptation across roadside deployments. Our system is integrated with CARLA and SUMO to create a high-fidelity DT that renders highway scenarios and captures traffic flows in real-time. Extensive experiments across diverse urban scenes show that FedMeta-GeoLane consistently outperforms baseline and meta-learning approaches, achieving lower geometric error and stronger generalization to unseen locations while drastically reducing communication overhead. This work lays the foundation for flexible, context-aware infrastructure modeling in DTs. The framework is publicly available at https://github.com/raynbowy23/FedMeta-GeoLane.git.
☆ HieraRS: A Hierarchical Segmentation Paradigm for Remote Sensing Enabling Multi-Granularity Interpretation and Cross-Domain Transfer
Hierarchical land cover and land use (LCLU) classification aims to assign pixel-wise labels with multiple levels of semantic granularity to remote sensing (RS) imagery. However, existing deep learning-based methods face two major challenges: 1) They predominantly adopt a flat classification paradigm, which limits their ability to generate end-to-end multi-granularity hierarchical predictions aligned with tree-structured hierarchies used in practice. 2) Most cross-domain studies focus on performance degradation caused by sensor or scene variations, with limited attention to transferring LCLU models to cross-domain tasks with heterogeneous hierarchies (e.g., LCLU to crop classification). These limitations hinder the flexibility and generalization of LCLU models in practical applications. To address these challenges, we propose HieraRS, a novel hierarchical interpretation paradigm that enables multi-granularity predictions and supports the efficient transfer of LCLU models to cross-domain tasks with heterogeneous tree-structured hierarchies. We introduce the Bidirectional Hierarchical Consistency Constraint Mechanism (BHCCM), which can be seamlessly integrated into mainstream flat classification models to generate hierarchical predictions, while improving both semantic consistency and classification accuracy. Furthermore, we present TransLU, a dual-branch cross-domain transfer framework comprising two key components: Cross-Domain Knowledge Sharing (CDKS) and Cross-Domain Semantic Alignment (CDSA). TransLU supports dynamic category expansion and facilitates the effective adaptation of LCLU models to heterogeneous hierarchies. In addition, we construct MM-5B, a large-scale multi-modal hierarchical land use dataset featuring pixel-wise annotations. The code and MM-5B dataset will be released at: https://github.com/AI-Tianlong/HieraRS.
comment: 17 pages, 11 figures
☆ Ensemble of Weak Spectral Total Variation Learners: a PET-CT Case Study
Solving computer vision problems through machine learning, one often encounters lack of sufficient training data. To mitigate this we propose the use of ensembles of weak learners based on spectral total-variation (STV) features (Gilboa 2014). The features are related to nonlinear eigenfunctions of the total-variation subgradient and can characterize well textures at various scales. It was shown (Burger et-al 2016) that, in the one-dimensional case, orthogonal features are generated, whereas in two-dimensions the features are empirically lowly correlated. Ensemble learning theory advocates the use of lowly correlated weak learners. We thus propose here to design ensembles using learners based on STV features. To show the effectiveness of this paradigm we examine a hard real-world medical imaging problem: the predictive value of computed tomography (CT) data for high uptake in positron emission tomography (PET) for patients suspected of skeletal metastases. The database consists of 457 scans with 1524 unique pairs of registered CT and PET slices. Our approach is compared to deep-learning methods and to Radiomics features, showing STV learners perform best (AUC=0.87), compared to neural nets (AUC=0.75) and Radiomics (AUC=0.79). We observe that fine STV scales in CT images are especially indicative for the presence of high uptake in PET.
☆ RoundaboutHD: High-Resolution Real-World Urban Environment Benchmark for Multi-Camera Vehicle Tracking
The multi-camera vehicle tracking (MCVT) framework holds significant potential for smart city applications, including anomaly detection, traffic density estimation, and suspect vehicle tracking. However, current publicly available datasets exhibit limitations, such as overly simplistic scenarios, low-resolution footage, and insufficiently diverse conditions, creating a considerable gap between academic research and real-world scenario. To fill this gap, we introduce RoundaboutHD, a comprehensive, high-resolution multi-camera vehicle tracking benchmark dataset specifically designed to represent real-world roundabout scenarios. RoundaboutHD provides a total of 40 minutes of labelled video footage captured by four non-overlapping, high-resolution (4K resolution, 15 fps) cameras. In total, 512 unique vehicle identities are annotated across different camera views, offering rich cross-camera association data. RoundaboutHD offers temporal consistency video footage and enhanced challenges, including increased occlusions and nonlinear movement inside the roundabout. In addition to the full MCVT dataset, several subsets are also available for object detection, single camera tracking, and image-based vehicle re-identification (ReID) tasks. Vehicle model information and camera modelling/ geometry information are also included to support further analysis. We provide baseline results for vehicle detection, single-camera tracking, image-based vehicle re-identification, and multi-camera tracking. The dataset and the evaluation code are publicly available at: https://github.com/siri-rouser/RoundaboutHD.git
☆ Learning human-to-robot handovers through 3D scene reconstruction
Learning robot manipulation policies from raw, real-world image data requires a large number of robot-action trials in the physical environment. Although training using simulations offers a cost-effective alternative, the visual domain gap between simulation and robot workspace remains a major limitation. Gaussian Splatting visual reconstruction methods have recently provided new directions for robot manipulation by generating realistic environments. In this paper, we propose the first method for learning supervised-based robot handovers solely from RGB images without the need of real-robot training or real-robot data collection. The proposed policy learner, Human-to-Robot Handover using Sparse-View Gaussian Splatting (H2RH-SGS), leverages sparse-view Gaussian Splatting reconstruction of human-to-robot handover scenes to generate robot demonstrations containing image-action pairs captured with a camera mounted on the robot gripper. As a result, the simulated camera pose changes in the reconstructed scene can be directly translated into gripper pose changes. We train a robot policy on demonstrations collected with 16 household objects and {\em directly} deploy this policy in the real environment. Experiments in both Gaussian Splatting reconstructed scene and real-world human-to-robot handover experiments demonstrate that H2RH-SGS serves as a new and effective representation for the human-to-robot handover task.
comment: 8 pages, 6 figures, 2 table
♻ ☆ X-Dancer: Expressive Music to Human Dance Video Generation ICCV 2025
We present X-Dancer, a novel zero-shot music-driven image animation pipeline that creates diverse and long-range lifelike human dance videos from a single static image. As its core, we introduce a unified transformer-diffusion framework, featuring an autoregressive transformer model that synthesize extended and music-synchronized token sequences for 2D body, head and hands poses, which then guide a diffusion model to produce coherent and realistic dance video frames. Unlike traditional methods that primarily generate human motion in 3D, X-Dancer addresses data limitations and enhances scalability by modeling a wide spectrum of 2D dance motions, capturing their nuanced alignment with musical beats through readily available monocular videos. To achieve this, we first build a spatially compositional token representation from 2D human pose labels associated with keypoint confidences, encoding both large articulated body movements (e.g., upper and lower body) and fine-grained motions (e.g., head and hands). We then design a music-to-motion transformer model that autoregressively generates music-aligned dance pose token sequences, incorporating global attention to both musical style and prior motion context. Finally we leverage a diffusion backbone to animate the reference image with these synthesized pose tokens through AdaIN, forming a fully differentiable end-to-end framework. Experimental results demonstrate that X-Dancer is able to produce both diverse and characterized dance videos, substantially outperforming state-of-the-art methods in term of diversity, expressiveness and realism. Code and model will be available for research purposes.
comment: ICCV 2025. Project Page: https://zeyuan-chen.com/X-Dancer/
♻ ☆ GoalNet: Goal Areas Oriented Pedestrian Trajectory Prediction
Predicting the future trajectories of pedestrians on the road is an important task for autonomous driving. The pedestrian trajectory prediction is affected by scene paths, pedestrian's intentions and decision-making, which is a multi-modal problem. Most recent studies use past trajectories to predict a variety of potential future trajectory distributions, which do not account for the scene context and pedestrian targets. Instead of predicting the future trajectory directly, we propose to use scene context and observed trajectory to predict the goal points first, and then reuse the goal points to predict the future trajectories. By leveraging the information from scene context and observed trajectory, the uncertainty can be limited to a few target areas, which represent the "goals" of the pedestrians. In this paper, we propose GoalNet, a new trajectory prediction neural network based on the goal areas of a pedestrian. Our network can predict both pedestrian's trajectories and bounding boxes. The overall model is efficient and modular, and its outputs can be changed according to the usage scenario. Experimental results show that GoalNet significantly improves the previous state-of-the-art performance by 48.7% on the JAAD and 40.8% on the PIE dataset.
♻ ☆ InsViE-1M: Effective Instruction-based Video Editing with Elaborate Dataset Construction ICCV 2025
Instruction-based video editing allows effective and interactive editing of videos using only instructions without extra inputs such as masks or attributes. However, collecting high-quality training triplets (source video, edited video, instruction) is a challenging task. Existing datasets mostly consist of low-resolution, short duration, and limited amount of source videos with unsatisfactory editing quality, limiting the performance of trained editing models. In this work, we present a high-quality Instruction-based Video Editing dataset with 1M triplets, namely InsViE-1M. We first curate high-resolution and high-quality source videos and images, then design an effective editing-filtering pipeline to construct high-quality editing triplets for model training. For a source video, we generate multiple edited samples of its first frame with different intensities of classifier-free guidance, which are automatically filtered by GPT-4o with carefully crafted guidelines. The edited first frame is propagated to subsequent frames to produce the edited video, followed by another round of filtering for frame quality and motion evaluation. We also generate and filter a variety of video editing triplets from high-quality images. With the InsViE-1M dataset, we propose a multi-stage learning strategy to train our InsViE model, progressively enhancing its instruction following and editing ability. Extensive experiments demonstrate the advantages of our InsViE-1M dataset and the trained model over state-of-the-art works. Codes are available at \href{https://github.com/langmanbusi/InsViE}{InsViE}.
comment: Accepted by ICCV 2025
♻ ☆ Synergistic Prompting for Robust Visual Recognition with Missing Modalities
Large-scale multi-modal models have demonstrated remarkable performance across various visual recognition tasks by leveraging extensive paired multi-modal training data. However, in real-world applications, the presence of missing or incomplete modality inputs often leads to significant performance degradation. Recent research has focused on prompt-based strategies to tackle this issue; however, existing methods are hindered by two major limitations: (1) static prompts lack the flexibility to adapt to varying missing-data conditions, and (2) basic prompt-tuning methods struggle to ensure reliable performance when critical modalities are missing.To address these challenges, we propose a novel Synergistic Prompting (SyP) framework for robust visual recognition with missing modalities. The proposed SyP introduces two key innovations: (I) a Dynamic Adapter, which computes adaptive scaling factors to dynamically generate prompts, replacing static parameters for flexible multi-modal adaptation, and (II) a Synergistic Prompting Strategy, which combines static and dynamic prompts to balance information across modalities, ensuring robust reasoning even when key modalities are missing. The proposed SyP achieves significant performance improvements over existing approaches across three widely-used visual recognition datasets, demonstrating robustness under diverse missing rates and conditions. Extensive experiments and ablation studies validate its effectiveness in handling missing modalities, highlighting its superior adaptability and reliability.
Information Retrieval 14
☆ Digital gazetteers: review and prospects for place name knowledge bases
Gazetteers typically store data on place names, place types and the associated coordinates. They play an essential role in disambiguating place names in online geographical information retrieval systems for navigation and mapping, detecting and disambiguating place names in text, and providing coordinates. Currently there are many gazetteers in use derived from many sources, with no commonly accepted standard for encoding the data. Most gazetteers are also very limited in the extent to which they represent the multiple facets of the named places yet they have potential to assist user search for locations with specific physical, commercial, social or cultural characteristics. With a view to understanding digital gazetteer technologies and advancing their future effectiveness for information retrieval, we provide a review of data sources, components, software and data management technologies, data quality and volunteered data, and methods for matching sources that refer to the same real-world places. We highlight the need for future work on richer representation of named places, the temporal evolution of place identity and location, and the development of more effective methods for data integration.
☆ Improving Korean-English Cross-Lingual Retrieval: A Data-Centric Study of Language Composition and Model Merging
With the increasing utilization of multilingual text information, Cross-Lingual Information Retrieval (CLIR) has become a crucial research area. However, the impact of training data composition on both CLIR and Mono-Lingual Information Retrieval (IR) performance remains under-explored. To systematically investigate this data-centric aspect, we construct linguistically parallel Korean-English datasets and train retrieval models with various language combinations. Our experiments reveal that the language composition of training data significantly influences IR performance, exhibiting important inter-lingual correlations: CLIR performance improves with specific language pairs, while Mono-Lingual IR performance declines. Our work demonstrates that Model Merging can effectively mitigate this trade-off, achieving strong CLIR results while preserving Mono-Lingual IR capabilities. Our findings underscore the effects of linguistic configuration of training data on both CLIR and Mono-Lingual IR, and present Model Merging as a viable strategy to optimize performance across these tasks.
☆ CUE-RAG: Towards Accurate and Cost-Efficient Graph-Based RAG via Multi-Partite Graph and Query-Driven Iterative Retrieval
Despite the remarkable progress of Large Language Models (LLMs), their performance in question answering (QA) remains limited by the lack of domain-specific and up-to-date knowledge. Retrieval-Augmented Generation (RAG) addresses this limitation by incorporating external information, often from graph-structured data. However, existing graph-based RAG methods suffer from poor graph quality due to incomplete extraction and insufficient utilization of query information during retrieval. To overcome these limitations, we propose CUE-RAG, a novel approach that introduces (1) a multi-partite graph index incorporates text Chunks, knowledge Units, and Entities to capture semantic content at multiple levels of granularity, (2) a hybrid extraction strategy that reduces LLM token usage while still producing accurate and disambiguated knowledge units, and (3) Q-Iter, a query-driven iterative retrieval strategy that enhances relevance through semantic search and constrained graph traversal. Experiments on three QA benchmarks show that CUE-RAG significantly outperforms state-of-the-art baselines, achieving up to 99.33% higher Accuracy and 113.51% higher F1 score while reducing indexing costs by 72.58%. Remarkably, CUE-RAG matches or outperforms baselines even without using an LLM for indexing. These results demonstrate the effectiveness and cost-efficiency of CUE-RAG in advancing graph-based RAG systems.
☆ DS@GT at LongEval: Evaluating Temporal Performance in Web Search Systems and Topics with Two-Stage Retrieval
Information Retrieval (IR) models are often trained on static datasets, making them vulnerable to performance degradation as web content evolves. The DS@GT competition team participated in the Longitudinal Evaluation of Model Performance (LongEval) lab at CLEF 2025, which evaluates IR systems across temporally distributed web snapshots. Our analysis of the Qwant web dataset includes exploratory data analysis with topic modeling over time. The two-phase retrieval system employs sparse keyword searches, utilizing query expansion and document reranking. Our best system achieves an average NDCG@10 of 0.296 across the entire training and test dataset, with an overall best score of 0.395 on 2023-05. The accompanying source code for this paper is at https://github.com/dsgt-arc/longeval-2025
☆ Distillation versus Contrastive Learning: How to Train Your Rerankers
Training text rerankers is crucial for information retrieval. Two primary strategies are widely used: contrastive learning (optimizing directly on ground-truth labels) and knowledge distillation (transferring knowledge from a larger reranker). While both have been studied in the literature, a clear comparison of their effectiveness for training cross-encoder rerankers under practical conditions is needed. This paper empirically compares these strategies by training rerankers of different sizes and architectures using both methods on the same data, with a strong contrastive learning model acting as the distillation teacher. Our results show that knowledge distillation generally yields better in-domain and out-of-domain ranking performance than contrastive learning when distilling from a larger teacher model. This finding is consistent across student model sizes and architectures. However, distilling from a teacher of the same capacity does not provide the same advantage, particularly for out-of-domain tasks. These findings offer practical guidance for choosing a training strategy based on available teacher models. Therefore, we recommend using knowledge distillation to train smaller rerankers if a larger, more powerful teacher is accessible; in its absence, contrastive learning provides a strong and more reliable alternative otherwise.
☆ Towards Efficient Quantity Retrieval from Text:an Approach via Description Parsing and Weak Supervision
Quantitative facts are continually generated by companies and governments, supporting data-driven decision-making. While common facts are structured, many long-tail quantitative facts remain buried in unstructured documents, making them difficult to access. We propose the task of Quantity Retrieval: given a description of a quantitative fact, the system returns the relevant value and supporting evidence. Understanding quantity semantics in context is essential for this task. We introduce a framework based on description parsing that converts text into structured (description, quantity) pairs for effective retrieval. To improve learning, we construct a large paraphrase dataset using weak supervision based on quantity co-occurrence. We evaluate our approach on a large corpus of financial annual reports and a newly annotated quantity description dataset. Our method significantly improves top-1 retrieval accuracy from 30.98 percent to 64.66 percent.
comment: Extended version of the paper accepted in DEXA 2025
☆ Transfer Learning and Mixup for Fine-Grained Few-Shot Fungi Classification
Accurate identification of fungi species presents a unique challenge in computer vision due to fine-grained inter-species variation and high intra-species variation. This paper presents our approach for the FungiCLEF 2025 competition, which focuses on few-shot fine-grained visual categorization (FGVC) using the FungiTastic Few-Shot dataset. Our team (DS@GT) experimented with multiple vision transformer models, data augmentation, weighted sampling, and incorporating textual information. We also explored generative AI models for zero-shot classification using structured prompting but found them to significantly underperform relative to vision-based models. Our final model outperformed both competition baselines and highlighted the effectiveness of domain specific pretraining and balanced sampling strategies. Our approach ranked 35/74 on the private test set in post-completion evaluation, this suggests additional work can be done on metadata selection and domain-adapted multi-modal learning. Our code is available at https://github.com/dsgt-arc/fungiclef-2025.
☆ Infinite Video Understanding
The rapid advancements in Large Language Models (LLMs) and their multimodal extensions (MLLMs) have ushered in remarkable progress in video understanding. However, a fundamental challenge persists: effectively processing and comprehending video content that extends beyond minutes or hours. While recent efforts like Video-XL-2 have demonstrated novel architectural solutions for extreme efficiency, and advancements in positional encoding such as HoPE and VideoRoPE++ aim to improve spatio-temporal understanding over extensive contexts, current state-of-the-art models still encounter significant computational and memory constraints when faced with the sheer volume of visual tokens from lengthy sequences. Furthermore, maintaining temporal coherence, tracking complex events, and preserving fine-grained details over extended periods remain formidable hurdles, despite progress in agentic reasoning systems like Deep Video Discovery. This position paper posits that a logical, albeit ambitious, next frontier for multimedia research is Infinite Video Understanding -- the capability for models to continuously process, understand, and reason about video data of arbitrary, potentially never-ending duration. We argue that framing Infinite Video Understanding as a blue-sky research objective provides a vital north star for the multimedia, and the wider AI, research communities, driving innovation in areas such as streaming architectures, persistent memory mechanisms, hierarchical and adaptive representations, event-centric reasoning, and novel evaluation paradigms. Drawing inspiration from recent work on long/ultra-long video understanding and several closely related fields, we outline the core challenges and key research directions towards achieving this transformative capability.
☆ Analysing Health Misinformation with Advanced Centrality Metrics in Online Social Networks
The rapid spread of health misinformation on online social networks (OSNs) during global crises such as the COVID-19 pandemic poses challenges to public health, social stability, and institutional trust. Centrality metrics have long been pivotal in understanding the dynamics of information flow, particularly in the context of health misinformation. However, the increasing complexity and dynamism of online networks, especially during crises, highlight the limitations of these traditional approaches. This study introduces and compares three novel centrality metrics: dynamic influence centrality (DIC), health misinformation vulnerability centrality (MVC), and propagation centrality (PC). These metrics incorporate temporal dynamics, susceptibility, and multilayered network interactions. Using the FibVID dataset, we compared traditional and novel metrics to identify influential nodes, propagation pathways, and misinformation influencers. Traditional metrics identified 29 influential nodes, while the new metrics uncovered 24 unique nodes, resulting in 42 combined nodes, an increase of 44.83%. Baseline interventions reduced health misinformation by 50%, while incorporating the new metrics increased this to 62.5%, an improvement of 25%. To evaluate the broader applicability of the proposed metrics, we validated our framework on a second dataset, Monant Medical Misinformation, which covers a diverse range of health misinformation discussions beyond COVID-19. The results confirmed that the advanced metrics generalised successfully, identifying distinct influential actors not captured by traditional methods. In general, the findings suggest that a combination of traditional and novel centrality measures offers a more robust and generalisable framework for understanding and mitigating the spread of health misinformation in different online network contexts.
comment: 10 Pages, 2 figures, 3 tables, journal article in PLOS Digital Health (2025)
☆ GraphRunner: A Multi-Stage Framework for Efficient and Accurate Graph-Based Retrieval
Conventional Retrieval Augmented Generation (RAG) approaches are common in text-based applications. However, they struggle with structured, interconnected datasets like knowledge graphs, where understanding underlying relationships is crucial for accurate retrieval. A common direction in graph-based retrieval employs iterative, rule-based traversal guided by Large Language Models (LLMs). Such existing iterative methods typically combine reasoning with single hop traversal at each step, making them vulnerable to LLM reasoning errors and hallucinations that ultimately hinder the retrieval of relevant information. To address these limitations, we propose GraphRunner, a novel graph-based retrieval framework that operates in three distinct stages: planning, verification, and execution. This introduces high-level traversal actions that enable multi-hop exploration in a single step. It also generates a holistic traversal plan, which is verified against the graph structure and pre-defined traversal actions, reducing reasoning errors and detecting hallucinations before execution. GraphRunner significantly reduces LLM reasoning errors and detects hallucinations through validation. Our evaluation using the GRBench dataset shows that GraphRunner consistently outperforms existing approaches, achieving 10-50% performance improvements over the strongest baseline while reducing inference cost by 3.0-12.9x and response generation time by 2.5-7.1x, making it significantly more robust and efficient for graph-based retrieval tasks.
♻ ☆ Drowning in Documents: Consequences of Scaling Reranker Inference SIGIR 2025
Rerankers, typically cross-encoders, are computationally intensive but are frequently used because they are widely assumed to outperform cheaper initial IR systems. We challenge this assumption by measuring reranker performance for full retrieval, not just re-scoring first-stage retrieval. To provide a more robust evaluation, we prioritize strong first-stage retrieval using modern dense embeddings and test rerankers on a variety of carefully chosen, challenging tasks, including internally curated datasets to avoid contamination, and out-of-domain ones. Our empirical results reveal a surprising trend: the best existing rerankers provide initial improvements when scoring progressively more documents, but their effectiveness gradually declines and can even degrade quality beyond a certain limit. We hope that our findings will spur future research to improve reranking.
comment: Accepted to ReNeuIR 2025 Workshop at SIGIR 2025 Conference
♻ ☆ REGEN: A Dataset and Benchmarks with Natural Language Critiques and Narratives
This paper introduces a novel dataset REGEN (Reviews Enhanced with GEnerative Narratives), designed to benchmark the conversational capabilities of recommender Large Language Models (LLMs), addressing the limitations of existing datasets that primarily focus on sequential item prediction. REGEN extends the Amazon Product Reviews dataset by inpainting two key natural language features: (1) user critiques, representing user "steering" queries that lead to the selection of a subsequent item, and (2) narratives, rich textual outputs associated with each recommended item taking into account prior context. The narratives include product endorsements, purchase explanations, and summaries of user preferences. Further, we establish an end-to-end modeling benchmark for the task of conversational recommendation, where models are trained to generate both recommendations and corresponding narratives conditioned on user history (items and critiques). For this joint task, we introduce a modeling framework LUMEN (LLM-based Unified Multi-task Model with Critiques, Recommendations, and Narratives) which uses an LLM as a backbone for critiquing, retrieval and generation. We also evaluate the dataset's quality using standard auto-rating techniques and benchmark it by training both traditional and LLM-based recommender models. Our results demonstrate that incorporating critiques enhances recommendation quality by enabling the recommender to learn language understanding and integrate it with recommendation signals. Furthermore, LLMs trained on our dataset effectively generate both recommendations and contextual narratives, achieving performance comparable to state-of-the-art recommenders and language models.
♻ ☆ Collaborative filtering based on nonnegative/binary matrix factorization
Collaborative filtering generates recommendations by exploiting user-item similarities based on rating data, which often contains numerous unrated items. This paper proposes a nonnegative/binary matrix factorization (NBMF) algorithm modified for collaborative filtering and demonstrates that utilizing a low-latency Ising machine in NBMF is advantageous in terms of computation time. While previous studies have primarily applied NBMF to dense data, such as images, this study applies a modified NBMF to sparse data. Results show the benefits of using a low-latency Ising machine to implement the proposed method.
comment: 12 pages, 8 figures
♻ ☆ Generative Retrieval and Alignment Model: A New Paradigm for E-commerce Retrieval WWW2025
Traditional sparse and dense retrieval methods struggle to leverage general world knowledge and often fail to capture the nuanced features of queries and products. With the advent of large language models (LLMs), industrial search systems have started to employ LLMs to generate identifiers for product retrieval. Commonly used identifiers include (1) static/semantic IDs and (2) product term sets. The first approach requires creating a product ID system from scratch, missing out on the world knowledge embedded within LLMs. While the second approach leverages this general knowledge, the significant difference in word distribution between queries and products means that product-based identifiers often do not align well with user search queries, leading to missed product recalls. Furthermore, when queries contain numerous attributes, these algorithms generate a large number of identifiers, making it difficult to assess their quality, which results in low overall recall efficiency. To address these challenges, this paper introduces a novel e-commerce retrieval paradigm: the Generative Retrieval and Alignment Model (GRAM). GRAM employs joint training on text information from both queries and products to generate shared text identifier codes, effectively bridging the gap between queries and products. This approach not only enhances the connection between queries and products but also improves inference efficiency. The model uses a co-alignment strategy to generate codes optimized for maximizing retrieval efficiency. Additionally, it introduces a query-product scoring mechanism to compare product values across different codes, further boosting retrieval efficiency. Extensive offline and online A/B testing demonstrates that GRAM significantly outperforms traditional models and the latest generative retrieval models, confirming its effectiveness and practicality.
comment: Accepted by WWW2025
Machine Learning 146
☆ The Non-Linear Representation Dilemma: Is Causal Abstraction Enough for Mechanistic Interpretability?
The concept of causal abstraction got recently popularised to demystify the opaque decision-making processes of machine learning models; in short, a neural network can be abstracted as a higher-level algorithm if there exists a function which allows us to map between them. Notably, most interpretability papers implement these maps as linear functions, motivated by the linear representation hypothesis: the idea that features are encoded linearly in a model's representations. However, this linearity constraint is not required by the definition of causal abstraction. In this work, we critically examine the concept of causal abstraction by considering arbitrarily powerful alignment maps. In particular, we prove that under reasonable assumptions, any neural network can be mapped to any algorithm, rendering this unrestricted notion of causal abstraction trivial and uninformative. We complement these theoretical findings with empirical evidence, demonstrating that it is possible to perfectly map models to algorithms even when these models are incapable of solving the actual task; e.g., on an experiment using randomly initialised language models, our alignment maps reach 100% interchange-intervention accuracy on the indirect object identification task. This raises the non-linear representation dilemma: if we lift the linearity constraint imposed to alignment maps in causal abstraction analyses, we are left with no principled way to balance the inherent trade-off between these maps' complexity and accuracy. Together, these results suggest an answer to our title's question: causal abstraction is not enough for mechanistic interpretability, as it becomes vacuous without assumptions about how models encode information. Studying the connection between this information-encoding assumption and causal abstraction should lead to exciting future work.
comment: 42 pages, 17 figures, code available in github.com/densutter/non-linear-representation-dilemma
☆ NeuralOS: Towards Simulating Operating Systems via Neural Generative Models
We introduce NeuralOS, a neural framework that simulates graphical user interfaces (GUIs) of operating systems by directly predicting screen frames in response to user inputs such as mouse movements, clicks, and keyboard events. NeuralOS combines a recurrent neural network (RNN), which tracks computer state, with a diffusion-based neural renderer that generates screen images. The model is trained on a large-scale dataset of Ubuntu XFCE recordings, which include both randomly generated interactions and realistic interactions produced by AI agents. Experiments show that NeuralOS successfully renders realistic GUI sequences, accurately captures mouse interactions, and reliably predicts state transitions like application launches. Although modeling fine-grained keyboard interactions precisely remains challenging, NeuralOS offers a step toward creating fully adaptive, generative neural interfaces for future human-computer interaction systems.
☆ Filter Equivariant Functions: A symmetric account of length-general extrapolation on lists
What should a function that extrapolates beyond known input/output examples look like? This is a tricky question to answer in general, as any function matching the outputs on those examples can in principle be a correct extrapolant. We argue that a "good" extrapolant should follow certain kinds of rules, and here we study a particularly appealing criterion for rule-following in list functions: that the function should behave predictably even when certain elements are removed. In functional programming, a standard way to express such removal operations is by using a filter function. Accordingly, our paper introduces a new semantic class of functions -- the filter equivariant functions. We show that this class contains interesting examples, prove some basic theorems about it, and relate it to the well-known class of map equivariant functions. We also present a geometric account of filter equivariants, showing how they correspond naturally to certain simplicial structures. Our highlight result is the amalgamation algorithm, which constructs any filter-equivariant function's output by first studying how it behaves on sublists of the input, in a way that extrapolates perfectly.
comment: 18 pages, 2 figures
☆ One Token to Fool LLM-as-a-Judge
Generative reward models (also known as LLMs-as-judges), which use large language models (LLMs) to evaluate answer quality, are increasingly adopted in reinforcement learning with verifiable rewards (RLVR). They are often preferred over rigid rule-based metrics, especially for complex reasoning tasks involving free-form outputs. In this paradigm, an LLM is typically prompted to compare a candidate answer against a ground-truth reference and assign a binary reward indicating correctness. Despite the seeming simplicity of this comparison task, we find that generative reward models exhibit surprising vulnerabilities to superficial manipulations: non-word symbols (e.g., ":" or ".") or reasoning openers like "Thought process:" and "Let's solve this problem step by step." can often lead to false positive rewards. We demonstrate that this weakness is widespread across LLMs, datasets, and prompt formats, posing a serious threat for core algorithmic paradigms that rely on generative reward models, such as rejection sampling, preference optimization, and RLVR. To mitigate this issue, we introduce a simple yet effective data augmentation strategy and train a new generative reward model with substantially improved robustness. Our findings highlight the urgent need for more reliable LLM-based evaluation methods. We release our robust, general-domain reward model and its synthetic training data at https://huggingface.co/sarosavo/Master-RM and https://huggingface.co/datasets/sarosavo/Master-RM.
☆ Optimistic Exploration for Risk-Averse Constrained Reinforcement Learning
Risk-averse Constrained Reinforcement Learning (RaCRL) aims to learn policies that minimise the likelihood of rare and catastrophic constraint violations caused by an environment's inherent randomness. In general, risk-aversion leads to conservative exploration of the environment which typically results in converging to sub-optimal policies that fail to adequately maximise reward or, in some cases, fail to achieve the goal. In this paper, we propose an exploration-based approach for RaCRL called Optimistic Risk-averse Actor Critic (ORAC), which constructs an exploratory policy by maximising a local upper confidence bound of the state-action reward value function whilst minimising a local lower confidence bound of the risk-averse state-action cost value function. Specifically, at each step, the weighting assigned to the cost value is increased or decreased if it exceeds or falls below the safety constraint value. This way the policy is encouraged to explore uncertain regions of the environment to discover high reward states whilst still satisfying the safety constraints. Our experimental results demonstrate that the ORAC approach prevents convergence to sub-optimal policies and improves significantly the reward-cost trade-off in various continuous control tasks such as Safety-Gymnasium and a complex building energy management environment CityLearn.
☆ Greedy Low-Rank Gradient Compression for Distributed Learning with Convergence Guarantees
Distributed optimization is pivotal for large-scale signal processing and machine learning, yet communication overhead remains a major bottleneck. Low-rank gradient compression, in which the transmitted gradients are approximated by low-rank matrices to reduce communication, offers a promising remedy. Existing methods typically adopt either randomized or greedy compression strategies: randomized approaches project gradients onto randomly chosen subspaces, introducing high variance and degrading empirical performance; greedy methods select the most informative subspaces, achieving strong empirical results but lacking convergence guarantees. To address this gap, we propose GreedyLore--the first Greedy Low-Rank gradient compression algorithm for distributed learning with rigorous convergence guarantees. GreedyLore incorporates error feedback to correct the bias introduced by greedy compression and introduces a semi-lazy subspace update that ensures the compression operator remains contractive throughout all iterations. With these techniques, we prove that GreedyLore achieves a convergence rate of $\mathcal{O}(\sigma/\sqrt{NT} + 1/T)$ under standard optimizers such as MSGD and Adam--marking the first linear speedup convergence rate for low-rank gradient compression. Extensive experiments are conducted to validate our theoretical findings.
comment: 18 pages, 5 figures
☆ BlockFFN: Towards End-Side Acceleration-Friendly Mixture-of-Experts with Chunk-Level Activation Sparsity
To alleviate the computational burden of large language models (LLMs), architectures with activation sparsity, represented by mixture-of-experts (MoE), have attracted increasing attention. However, the non-differentiable and inflexible routing of vanilla MoE hurts model performance. Moreover, while each token activates only a few parameters, these sparsely-activated architectures exhibit low chunk-level sparsity, indicating that the union of multiple consecutive tokens activates a large ratio of parameters. Such a sparsity pattern is unfriendly for acceleration under low-resource conditions (e.g., end-side devices) and incompatible with mainstream acceleration techniques (e.g., speculative decoding). To address these challenges, we introduce a novel MoE architecture, BlockFFN, as well as its efficient training and deployment techniques. Specifically, we use a router integrating ReLU activation and RMSNorm for differentiable and flexible routing. Next, to promote both token-level sparsity (TLS) and chunk-level sparsity (CLS), CLS-aware training objectives are designed, making BlockFFN more acceleration-friendly. Finally, we implement efficient acceleration kernels, combining activation sparsity and speculative decoding for the first time. The experimental results demonstrate the superior performance of BlockFFN over other MoE baselines, achieving over 80% TLS and 70% 8-token CLS. Our kernels achieve up to 3.67$\times$ speedup on real end-side devices than dense models. All codes and checkpoints are available publicly (https://github.com/thunlp/BlockFFN).
comment: 21 pages, 7 figures, 15 tables
☆ A Hybrid Multi-Well Hopfield-CNN with Feature Extraction and K-Means for MNIST Classification
This study presents a hybrid model for classifying handwritten digits in the MNIST dataset, combining convolutional neural networks (CNNs) with a multi-well Hopfield network. The approach employs a CNN to extract high-dimensional features from input images, which are then clustered into class-specific prototypes using k-means clustering. These prototypes serve as attractors in a multi-well energy landscape, where a Hopfield network performs classification by minimizing an energy function that balances feature similarity and class assignment.The model's design enables robust handling of intraclass variability, such as diverse handwriting styles, while providing an interpretable framework through its energy-based decision process. Through systematic optimization of the CNN architecture and the number of wells, the model achieves a high test accuracy of 99.2% on 10,000 MNIST images, demonstrating its effectiveness for image classification tasks. The findings highlight the critical role of deep feature extraction and sufficient prototype coverage in achieving high performance, with potential for broader applications in pattern recognition.
☆ Penalizing Infeasible Actions and Reward Scaling in Reinforcement Learning with Offline Data ICML2025
Reinforcement learning with offline data suffers from Q-value extrapolation errors. To address this issue, we first demonstrate that linear extrapolation of the Q-function beyond the data range is particularly problematic. To mitigate this, we propose guiding the gradual decrease of Q-values outside the data range, which is achieved through reward scaling with layer normalization (RS-LN) and a penalization mechanism for infeasible actions (PA). By combining RS-LN and PA, we develop a new algorithm called PARS. We evaluate PARS across a range of tasks, demonstrating superior performance compared to state-of-the-art algorithms in both offline training and online fine-tuning on the D4RL benchmark, with notable success in the challenging AntMaze Ultra task.
comment: Accepted to ICML2025
☆ ML-Based Automata Simplification for Symbolic Accelerators
Symbolic accelerators are increasingly used for symbolic data processing in domains such as genomics, NLP, and cybersecurity. However, these accelerators face scalability issues due to excessive memory use and routing complexity, especially when targeting a large set. We present AutoSlim, a machine learning-based graph simplification framework designed to reduce the complexity of symbolic accelerators built on Non-deterministic Finite Automata (NFA) deployed on FPGA-based overlays such as NAPOLY+. AutoSlim uses Random Forest classification to prune low-impact transitions based on edge scores and structural features, significantly reducing automata graph density while preserving semantic correctness. Unlike prior tools, AutoSlim targets automated score-aware simplification with weighted transitions, enabling efficient ranking-based sequence analysis. We evaluated data sets (1K to 64K nodes) in NAPOLY+ and conducted performance measurements including latency, throughput, and resource usage. AutoSlim achieves up to 40 percent reduction in FPGA LUTs and over 30 percent pruning in transitions, while scaling to graphs an order of magnitude larger than existing benchmarks. Our results also demonstrate how hardware interconnection (fanout) heavily influences hardware cost and that AutoSlim's pruning mitigates resource blowup.
☆ Modeling Partially Observed Nonlinear Dynamical Systems and Efficient Data Assimilation via Discrete-Time Conditional Gaussian Koopman Network
A discrete-time conditional Gaussian Koopman network (CGKN) is developed in this work to learn surrogate models that can perform efficient state forecast and data assimilation (DA) for high-dimensional complex dynamical systems, e.g., systems governed by nonlinear partial differential equations (PDEs). Focusing on nonlinear partially observed systems that are common in many engineering and earth science applications, this work exploits Koopman embedding to discover a proper latent representation of the unobserved system states, such that the dynamics of the latent states are conditional linear, i.e., linear with the given observed system states. The modeled system of the observed and latent states then becomes a conditional Gaussian system, for which the posterior distribution of the latent states is Gaussian and can be efficiently evaluated via analytical formulae. The analytical formulae of DA facilitate the incorporation of DA performance into the learning process of the modeled system, which leads to a framework that unifies scientific machine learning (SciML) and data assimilation. The performance of discrete-time CGKN is demonstrated on several canonical problems governed by nonlinear PDEs with intermittency and turbulent features, including the viscous Burgers' equation, the Kuramoto-Sivashinsky equation, and the 2-D Navier-Stokes equations, with which we show that the discrete-time CGKN framework achieves comparable performance as the state-of-the-art SciML methods in state forecast and provides efficient and accurate DA results. The discrete-time CGKN framework also serves as an example to illustrate unifying the development of SciML models and their other outer-loop applications such as design optimization, inverse problems, and optimal control.
☆ Partitioned Hybrid Quantum Fourier Neural Operators for Scientific Quantum Machine Learning
We introduce the Partitioned Hybrid Quantum Fourier Neural Operator (PHQFNO), a generalization of the Quantum Fourier Neural Operator (QFNO) for scientific machine learning. PHQFNO partitions the Fourier operator computation across classical and quantum resources, enabling tunable quantum-classical hybridization and distributed execution across quantum and classical devices. The method extends QFNOs to higher dimensions and incorporates a message-passing framework to distribute data across different partitions. Input data are encoded into quantum states using unary encoding, and quantum circuit parameters are optimized using a variational scheme. We implement PHQFNO using PennyLane with PyTorch integration and evaluate it on Burgers' equation, incompressible and compressible Navier-Stokes equations. We show that PHQFNO recovers classical FNO accuracy. On incompressible Navier-Stokes, PHQFNO achieves higher accuracy than its classical counterparts. Finally, we perform a sensitivity analysis under input noise, confirming improved stability of PHQFNO over classical baselines.
☆ Hashing for Fast Pattern Set Selection ECML-PKDD 2025
Pattern set mining, which is the task of finding a good set of patterns instead of all patterns, is a fundamental problem in data mining. Many different definitions of what constitutes a good set have been proposed in recent years. In this paper, we consider the reconstruction error as a proxy measure for the goodness of the set, and concentrate on the adjacent problem of how to find a good set efficiently. We propose a method based on bottom-k hashing for efficiently selecting the set and extend the method for the common case where the patterns might only appear in approximate form in the data. Our approach has applications in tiling databases, Boolean matrix factorization, and redescription mining, among others. We show that our hashing-based approach is significantly faster than the standard greedy algorithm while obtaining almost equally good results in both synthetic and real-world data sets.
comment: 17 pages, 5 figures, to appear at ECML-PKDD 2025
☆ Adaptive Nonlinear Vector Autoregression: Robust Forecasting for Noisy Chaotic Time Series
Nonlinear vector autoregression (NVAR) and reservoir computing (RC) have shown promise in forecasting chaotic dynamical systems, such as the Lorenz-63 model and El Nino-Southern Oscillation. However, their reliance on fixed nonlinearities - polynomial expansions in NVAR or random feature maps in RC - limits their adaptability to high noise or real-world data. These methods also scale poorly in high-dimensional settings due to costly matrix inversion during readout computation. We propose an adaptive NVAR model that combines delay-embedded linear inputs with features generated by a shallow, learnable multi-layer perceptron (MLP). The MLP and linear readout are jointly trained using gradient-based optimization, enabling the model to learn data-driven nonlinearities while preserving a simple readout structure. Unlike standard NVAR, our approach avoids the need for an exhaustive and sensitive grid search over ridge and delay parameters. Instead, tuning is restricted to neural network hyperparameters, improving scalability. Initial experiments on chaotic systems tested under noise-free and synthetically noisy conditions showed that the adaptive model outperformed the standard NVAR in predictive accuracy and showed robust forecasting under noisy conditions with a lower observation frequency.
comment: 15 pages, 10 figures
☆ Catastrophic Forgetting Mitigation Through Plateau Phase Activity Profiling
Catastrophic forgetting in deep neural networks occurs when learning new tasks degrades performance on previously learned tasks due to knowledge overwriting. Among the approaches to mitigate this issue, regularization techniques aim to identify and constrain "important" parameters to preserve previous knowledge. In the highly nonconvex optimization landscape of deep learning, we propose a novel perspective: tracking parameters during the final training plateau is more effective than monitoring them throughout the entire training process. We argue that parameters that exhibit higher activity (movement and variability) during this plateau reveal directions in the loss landscape that are relatively flat, making them suitable for adaptation to new tasks while preserving knowledge from previous ones. Our comprehensive experiments demonstrate that this approach achieves superior performance in balancing catastrophic forgetting mitigation with strong performance on newly learned tasks.
☆ Monitoring Risks in Test-Time Adaptation
Encountering shifted data at test time is a ubiquitous challenge when deploying predictive models. Test-time adaptation (TTA) methods address this issue by continuously adapting a deployed model using only unlabeled test data. While TTA can extend the model's lifespan, it is only a temporary solution. Eventually the model might degrade to the point that it must be taken offline and retrained. To detect such points of ultimate failure, we propose pairing TTA with risk monitoring frameworks that track predictive performance and raise alerts when predefined performance criteria are violated. Specifically, we extend existing monitoring tools based on sequential testing with confidence sequences to accommodate scenarios in which the model is updated at test time and no test labels are available to estimate the performance metrics of interest. Our extensions unlock the application of rigorous statistical risk monitoring to TTA, and we demonstrate the effectiveness of our proposed TTA monitoring framework across a representative set of datasets, distribution shift types, and TTA methods.
☆ On the Effect of Regularization in Policy Mirror Descent
Policy Mirror Descent (PMD) has emerged as a unifying framework in reinforcement learning (RL) by linking policy gradient methods with a first-order optimization method known as mirror descent. At its core, PMD incorporates two key regularization components: (i) a distance term that enforces a trust region for stable policy updates and (ii) an MDP regularizer that augments the reward function to promote structure and robustness. While PMD has been extensively studied in theory, empirical investigations remain scarce. This work provides a large-scale empirical analysis of the interplay between these two regularization techniques, running over 500k training seeds on small RL environments. Our results demonstrate that, although the two regularizers can partially substitute each other, their precise combination is critical for achieving robust performance. These findings highlight the potential for advancing research on more robust algorithms in RL, particularly with respect to hyperparameter sensitivity.
comment: Accepted at RLC
☆ SPLASH! Sample-efficient Preference-based inverse reinforcement learning for Long-horizon Adversarial tasks from Suboptimal Hierarchical demonstrations
Inverse Reinforcement Learning (IRL) presents a powerful paradigm for learning complex robotic tasks from human demonstrations. However, most approaches make the assumption that expert demonstrations are available, which is often not the case. Those that allow for suboptimality in the demonstrations are not designed for long-horizon goals or adversarial tasks. Many desirable robot capabilities fall into one or both of these categories, thus highlighting a critical shortcoming in the ability of IRL to produce field-ready robotic agents. We introduce Sample-efficient Preference-based inverse reinforcement learning for Long-horizon Adversarial tasks from Suboptimal Hierarchical demonstrations (SPLASH), which advances the state-of-the-art in learning from suboptimal demonstrations to long-horizon and adversarial settings. We empirically validate SPLASH on a maritime capture-the-flag task in simulation, and demonstrate real-world applicability with sim-to-real translation experiments on autonomous unmanned surface vehicles. We show that our proposed methods allow SPLASH to significantly outperform the state-of-the-art in reward learning from suboptimal demonstrations.
☆ Domain-Informed Operation Excellence of Gas Turbine System with Machine Learning
The domain-consistent adoption of artificial intelligence (AI) remains low in thermal power plants due to the black-box nature of AI algorithms and low representation of domain knowledge in conventional data-centric analytics. In this paper, we develop a MAhalanobis Distance-based OPTimization (MAD-OPT) framework that incorporates the Mahalanobis distance-based constraint to introduce domain knowledge into data-centric analytics. The developed MAD-OPT framework is applied to maximize thermal efficiency and minimize turbine heat rate for a 395 MW capacity gas turbine system. We demonstrate that the MAD-OPT framework can estimate domain-informed optimal process conditions under different ambient conditions, and the optimal solutions are found to be robust as evaluated by Monte Carlo simulations. We also apply the MAD-OPT framework to estimate optimal process conditions beyond the design power generation limit of the gas turbine system, and have found comparable results with the actual data of the power plant. We demonstrate that implementing data-centric optimization analytics without incorporating domain-informed constraints may provide ineffective solutions that may not be implementable in the real operation of the gas turbine system. This research advances the integration of the data-driven domain knowledge into machine learning-powered analytics that enhances the domain-informed operation excellence and paves the way for safe AI adoption in thermal power systems.
☆ Forget Me Not: Fighting Local Overfitting with Knowledge Fusion and Distillation
Overfitting in deep neural networks occurs less frequently than expected. This is a puzzling observation, as theory predicts that greater model capacity should eventually lead to overfitting -- yet this is rarely seen in practice. But what if overfitting does occur, not globally, but in specific sub-regions of the data space? In this work, we introduce a novel score that measures the forgetting rate of deep models on validation data, capturing what we term local overfitting: a performance degradation confined to certain regions of the input space. We demonstrate that local overfitting can arise even without conventional overfitting, and is closely linked to the double descent phenomenon. Building on these insights, we introduce a two-stage approach that leverages the training history of a single model to recover and retain forgotten knowledge: first, by aggregating checkpoints into an ensemble, and then by distilling it into a single model of the original size, thus enhancing performance without added inference cost. Extensive experiments across multiple datasets, modern architectures, and training regimes validate the effectiveness of our approach. Notably, in the presence of label noise, our method -- Knowledge Fusion followed by Knowledge Distillation -- outperforms both the original model and independently trained ensembles, achieving a rare win-win scenario: reduced training and inference complexity.
comment: arXiv admin note: substantial text overlap with arXiv:2412.12968
☆ The Impact of Automatic Speech Transcription on Speaker Attribution
Speaker attribution from speech transcripts is the task of identifying a speaker from the transcript of their speech based on patterns in their language use. This task is especially useful when the audio is unavailable (e.g. deleted) or unreliable (e.g. anonymized speech). Prior work in this area has primarily focused on the feasibility of attributing speakers using transcripts produced by human annotators. However, in real-world settings, one often only has more errorful transcripts produced by automatic speech recognition (ASR) systems. In this paper, we conduct what is, to our knowledge, the first comprehensive study of the impact of automatic transcription on speaker attribution performance. In particular, we study the extent to which speaker attribution performance degrades in the face of transcription errors, as well as how properties of the ASR system impact attribution. We find that attribution is surprisingly resilient to word-level transcription errors and that the objective of recovering the true transcript is minimally correlated with attribution performance. Overall, our findings suggest that speaker attribution on more errorful transcripts produced by ASR is as good, if not better, than attribution based on human-transcribed data, possibly because ASR transcription errors can capture speaker-specific features revealing of speaker identity.
☆ Safe Deep Reinforcement Learning for Resource Allocation with Peak Age of Information Violation Guarantees
In Wireless Networked Control Systems (WNCSs), control and communication systems must be co-designed due to their strong interdependence. This paper presents a novel optimization theory-based safe deep reinforcement learning (DRL) framework for ultra-reliable WNCSs, ensuring constraint satisfaction while optimizing performance, for the first time in the literature. The approach minimizes power consumption under key constraints, including Peak Age of Information (PAoI) violation probability, transmit power, and schedulability in the finite blocklength regime. PAoI violation probability is uniquely derived by combining stochastic maximum allowable transfer interval (MATI) and maximum allowable packet delay (MAD) constraints in a multi-sensor network. The framework consists of two stages: optimization theory and safe DRL. The first stage derives optimality conditions to establish mathematical relationships among variables, simplifying and decomposing the problem. The second stage employs a safe DRL model where a teacher-student framework guides the DRL agent (student). The control mechanism (teacher) evaluates compliance with system constraints and suggests the nearest feasible action when needed. Extensive simulations show that the proposed framework outperforms rule-based and other optimization theory based DRL benchmarks, achieving faster convergence, higher rewards, and greater stability.
comment: 15 Pages, to be published in IEEE Transactions on Communications
☆ Scaling Attention to Very Long Sequences in Linear Time with Wavelet-Enhanced Random Spectral Attention (WERSA)
Transformer models are computationally costly on long sequences since regular attention has quadratic $O(n^2)$ time complexity. We introduce Wavelet-Enhanced Random Spectral Attention (WERSA), a novel mechanism of linear $O(n)$ time complexity that is pivotal to enable successful long-sequence processing without the performance trade-off. WERSA merges content-adaptive random spectral features together with multi-resolution Haar wavelets and learnable parameters to selectively attend to informative scales of data while preserving linear efficiency. Large-scale comparisons \textbf{on single GPU} and across various benchmarks (vision, NLP, hierarchical reasoning) and various attention mechanisms (like Multiheaded Attention, Flash-Attention-2, FNet, Linformer, Performer, Waveformer), reveal uniform advantages of WERSA. It achieves best accuracy in all tests. On ArXiv classification, WERSA improves accuracy over vanilla attention by 1.2\% (86.2\% vs 85.0\%) while cutting training time by 81\% (296s vs 1554s) and FLOPS by 73.4\% (26.2G vs 98.4G). Significantly, WERSA excels where vanilla and FlashAttention-2 fail: on ArXiv-128k's extremely lengthy sequences, it achieves best accuracy (79.1\%) and AUC (0.979) among viable methods, operating on data that gives Out-Of-Memory errors to quadratic methods while being \textbf{twice as fast} as Waveformer, its next-best competitor. By significantly reducing computational loads without compromising accuracy, WERSA makes possible more practical, more affordable, long-context models, in particular on low-resource hardware, for more sustainable and more scalable AI development.
comment: 10 pages, 1 figure
☆ Entangled Threats: A Unified Kill Chain Model for Quantum Machine Learning Security
Quantum Machine Learning (QML) systems inherit vulnerabilities from classical machine learning while introducing new attack surfaces rooted in the physical and algorithmic layers of quantum computing. Despite a growing body of research on individual attack vectors - ranging from adversarial poisoning and evasion to circuit-level backdoors, side-channel leakage, and model extraction - these threats are often analyzed in isolation, with unrealistic assumptions about attacker capabilities and system environments. This fragmentation hampers the development of effective, holistic defense strategies. In this work, we argue that QML security requires more structured modeling of the attack surface, capturing not only individual techniques but also their relationships, prerequisites, and potential impact across the QML pipeline. We propose adapting kill chain models, widely used in classical IT and cybersecurity, to the quantum machine learning context. Such models allow for structured reasoning about attacker objectives, capabilities, and possible multi-stage attack paths - spanning reconnaissance, initial access, manipulation, persistence, and exfiltration. Based on extensive literature analysis, we present a detailed taxonomy of QML attack vectors mapped to corresponding stages in a quantum-aware kill chain framework that is inspired by the MITRE ATLAS for classical machine learning. We highlight interdependencies between physical-level threats (like side-channel leakage and crosstalk faults), data and algorithm manipulation (such as poisoning or circuit backdoors), and privacy attacks (including model extraction and training data inference). This work provides a foundation for more realistic threat modeling and proactive security-in-depth design in the emerging field of quantum machine learning.
comment: Accepted for publication at IEEE International Conference on Quantum Computing and Engineering (QCE) 2025
☆ Towards Collaborative Fairness in Federated Learning Under Imbalanced Covariate Shift KDD
Collaborative fairness is a crucial challenge in federated learning. However, existing approaches often overlook a practical yet complex form of heterogeneity: imbalanced covariate shift. We provide a theoretical analysis of this setting, which motivates the design of FedAKD (Federated Asynchronous Knowledge Distillation)- simple yet effective approach that balances accurate prediction with collaborative fairness. FedAKD consists of client and server updates. In the client update, we introduce a novel asynchronous knowledge distillation strategy based on our preliminary analysis, which reveals that while correctly predicted samples exhibit similar feature distributions across clients, incorrectly predicted samples show significant variability. This suggests that imbalanced covariate shift primarily arises from misclassified samples. Leveraging this insight, our approach first applies traditional knowledge distillation to update client models while keeping the global model fixed. Next, we select correctly predicted high-confidence samples and update the global model using these samples while keeping client models fixed. The server update simply aggregates all client models. We further provide a theoretical proof of FedAKD's convergence. Experimental results on public datasets (FashionMNIST and CIFAR10) and a real-world Electronic Health Records (EHR) dataset demonstrate that FedAKD significantly improves collaborative fairness, enhances predictive accuracy, and fosters client participation even under highly heterogeneous data distributions.
comment: 18 pages, accepted to the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD' 25), Toronto, Canada, August 3-7 2025
☆ AgentsNet: Coordination and Collaborative Reasoning in Multi-Agent LLMs
Large-language models (LLMs) have demonstrated powerful problem-solving capabilities, in particular when organized in multi-agent systems. However, the advent of such systems also raises several questions on the ability of a complex network of agents to effectively self-organize and collaborate. While measuring performance on standard reasoning benchmarks indicates how well multi-agent systems can solve reasoning tasks, it is unclear whether these systems are able to leverage their topology effectively. Here, we propose AgentsNet, a new benchmark for multi-agent reasoning. By drawing inspiration from classical problems in distributed systems and graph theory, AgentsNet measures the ability of multi-agent systems to collaboratively form strategies for problem-solving, self-organization, and effective communication given a network topology. We evaluate a variety of baseline methods on AgentsNet including homogeneous networks of agents which first have to agree on basic protocols for organization and communication. We find that some frontier LLMs are already demonstrating strong performance for small networks but begin to fall off once the size of the network scales. While existing multi-agent benchmarks cover at most 2-5 agents, AgentsNet is practically unlimited in size and can scale with new generations of LLMs. As such, we also probe frontier models in a setup with up to 100 agents.
comment: Preprint
☆ Emergent Natural Language with Communication Games for Improving Image Captioning Capabilities without Additional Data
Image captioning is an important problem in developing various AI systems, and these tasks require large volumes of annotated images to train the models. Since all existing labelled datasets are already used for training the large Vision Language Models (VLMs), it becomes challenging to improve the performance of the same. Considering this, it is essential to consider the unsupervised image captioning performance, which remains relatively under-explored. To that end, we propose LoGIC (Lewis Communication Game for Image Captioning), a Multi-agent Reinforcement Learning game. The proposed method consists of two agents, a 'speaker' and a 'listener', with the objective of learning a strategy for communicating in natural language. We train agents in the cooperative common-reward setting using the GRPO algorithm and show that improvement in image captioning performance emerges as a consequence of the agents learning to play the game. We show that using pre-trained VLMs as the 'speaker' and Large Language Model (LLM) for language understanding in the 'listener', we achieved a $46$ BLEU score after fine-tuning using LoGIC without additional labels, a $2$ units advantage in absolute metrics compared to the $44$ BLEU score of the vanilla VLM. Additionally, we replace the VLM from the 'speaker' with lightweight components: (i) a ViT for image perception and (ii) a GPT2 language generation, and train them from scratch using LoGIC, obtaining a $31$ BLEU score in the unsupervised setting, a $10$ points advantage over existing unsupervised image-captioning methods.
☆ Remote Sensing Reveals Adoption of Sustainable Rice Farming Practices Across Punjab, India
Rice cultivation consumes 24-30% of global freshwater, creating critical water management challenges in major rice-producing regions. Sustainable irrigation practices like direct seeded rice (DSR) and alternate wetting and drying (AWD) can reduce water use by 20-40% while maintaining yields, helping secure long-term agricultural productivity as water scarcity intensifies - a key component of the Zero Hunger Sustainable Development Goal. However, limited data on adoption rates of these practices prevents evidence-based policymaking and targeted resource allocation. We developed a novel remote sensing framework to monitor sustainable water management practices at scale in Punjab, India - a region facing severe groundwater depletion of 41.6 cm/year. To collect essential ground truth data, we partnered with the Nature Conservancy's Promoting Regenerative and No-burn Agriculture (PRANA) program, which trained approximately 1,400 farmers on water-saving techniques while documenting their field-level practices. Using this data, we created a classification system with Sentinel-1 satellite imagery that separates water management along sowing and irrigation dimensions. Our approach achieved a 78% F1-score in distinguishing DSR from traditional puddled transplanted rice without requiring prior knowledge of planting dates. We demonstrated scalability by mapping DSR adoption across approximately 3 million agricultural plots in Punjab, with district-level predictions showing strong correlation (Pearson=0.77, RBO= 0.77) with government records. This study provides policymakers with a powerful tool to track sustainable water management adoption, target interventions, and measure program impacts at scale.
comment: Dataset and code will be published shortly and links updated in v2
☆ ADAPT: A Pseudo-labeling Approach to Combat Concept Drift in Malware Detection
Machine learning models are commonly used for malware classification; however, they suffer from performance degradation over time due to concept drift. Adapting these models to changing data distributions requires frequent updates, which rely on costly ground truth annotations. While active learning can reduce the annotation burden, leveraging unlabeled data through semi-supervised learning remains a relatively underexplored approach in the context of malware detection. In this research, we introduce \texttt{ADAPT}, a novel pseudo-labeling semi-supervised algorithm for addressing concept drift. Our model-agnostic method can be applied to various machine learning models, including neural networks and tree-based algorithms. We conduct extensive experiments on five diverse malware detection datasets spanning Android, Windows, and PDF domains. The results demonstrate that our method consistently outperforms baseline models and competitive benchmarks. This work paves the way for more effective adaptation of machine learning models to concept drift in malware detection.
☆ AbbIE: Autoregressive Block-Based Iterative Encoder for Efficient Sequence Modeling NeurIPS 2025
We introduce the Autoregressive Block-Based Iterative Encoder (AbbIE), a novel recursive generalization of the encoder-only Transformer architecture, which achieves better perplexity than a standard Transformer and allows for the dynamic scaling of compute resources at test time. This simple, recursive approach is a complement to scaling large language model (LLM) performance through parameter and token counts. AbbIE performs its iterations in latent space, but unlike latent reasoning models, does not require a specialized dataset or training protocol. We show that AbbIE upward generalizes (ability to generalize to arbitrary iteration lengths) at test time by only using 2 iterations during train time, far outperforming alternative iterative methods. AbbIE's ability to scale its computational expenditure based on the complexity of the task gives it an up to \textbf{12\%} improvement in zero-shot in-context learning tasks versus other iterative and standard methods and up to 5\% improvement in language perplexity. The results from this study open a new avenue to Transformer performance scaling. We perform all of our evaluations on model sizes up to 350M parameters.
comment: 14 pages and 6 figures. Submitted to NeurIPS 2025
☆ STRAP: Spatial-Temporal Risk-Attentive Vehicle Trajectory Prediction for Autonomous Driving SC 2025
Accurate vehicle trajectory prediction is essential for ensuring safety and efficiency in fully autonomous driving systems. While existing methods primarily focus on modeling observed motion patterns and interactions with other vehicles, they often neglect the potential risks posed by the uncertain or aggressive behaviors of surrounding vehicles. In this paper, we propose a novel spatial-temporal risk-attentive trajectory prediction framework that incorporates a risk potential field to assess perceived risks arising from behaviors of nearby vehicles. The framework leverages a spatial-temporal encoder and a risk-attentive feature fusion decoder to embed the risk potential field into the extracted spatial-temporal feature representations for trajectory prediction. A risk-scaled loss function is further designed to improve the prediction accuracy of high-risk scenarios, such as short relative spacing. Experiments on the widely used NGSIM and HighD datasets demonstrate that our method reduces average prediction errors by 4.8% and 31.2% respectively compared to state-of-the-art approaches, especially in high-risk scenarios. The proposed framework provides interpretable, risk-aware predictions, contributing to more robust decision-making for autonomous driving systems.
comment: 6 pages, 3 figures, accepted at ITSC 2025
☆ SAM2RL: Towards Reinforcement Learning Memory Control in Segment Anything Model 2
Segment Anything Model 2 (SAM 2) has demonstrated strong performance in object segmentation tasks and has become the state-of-the-art for visual object tracking. The model stores information from previous frames in a memory bank, enabling temporal consistency across video sequences. Recent methods augment SAM 2 with hand-crafted update rules to better handle distractors, occlusions, and object motion. We propose a fundamentally different approach using reinforcement learning for optimizing memory updates in SAM 2 by framing memory control as a sequential decision-making problem. In an overfitting setup with a separate agent per video, our method achieves a relative improvement over SAM 2 that exceeds by more than three times the gains of existing heuristics. These results reveal the untapped potential of the memory bank and highlight reinforcement learning as a powerful alternative to hand-crafted update rules for memory control in visual object tracking.
☆ Quantum Algorithms for Projection-Free Sparse Convex Optimization
This paper considers the projection-free sparse convex optimization problem for the vector domain and the matrix domain, which covers a large number of important applications in machine learning and data science. For the vector domain $\mathcal{D} \subset \mathbb{R}^d$, we propose two quantum algorithms for sparse constraints that finds a $\varepsilon$-optimal solution with the query complexity of $O(\sqrt{d}/\varepsilon)$ and $O(1/\varepsilon)$ by using the function value oracle, reducing a factor of $O(\sqrt{d})$ and $O(d)$ over the best classical algorithm, respectively, where $d$ is the dimension. For the matrix domain $\mathcal{D} \subset \mathbb{R}^{d\times d}$, we propose two quantum algorithms for nuclear norm constraints that improve the time complexity to $\tilde{O}(rd/\varepsilon^2)$ and $\tilde{O}(\sqrt{r}d/\varepsilon^3)$ for computing the update step, reducing at least a factor of $O(\sqrt{d})$ over the best classical algorithm, where $r$ is the rank of the gradient matrix. Our algorithms show quantum advantages in projection-free sparse convex optimization problems as they outperform the optimal classical methods in dependence on the dimension $d$.
☆ CircFormerMoE: An End-to-End Deep Learning Framework for Circular RNA Splice Site Detection and Pairing in Plant Genomes
Circular RNAs (circRNAs) are important components of the non-coding RNA regulatory network. Previous circRNA identification primarily relies on high-throughput RNA sequencing (RNA-seq) data combined with alignment-based algorithms that detect back-splicing signals. However, these methods face several limitations: they can't predict circRNAs directly from genomic DNA sequences and relies heavily on RNA experimental data; they involve high computational costs due to complex alignment and filtering steps; and they are inefficient for large-scale or genome-wide circRNA prediction. The challenge is even greater in plants, where plant circRNA splice sites often lack the canonical GT-AG motif seen in human mRNA splicing, and no efficient deep learning model with strong generalization capability currently exists. Furthermore, the number of currently identified plant circRNAs is likely far lower than their true abundance. In this paper, we propose a deep learning framework named CircFormerMoE based on transformers and mixture-of experts for predicting circRNAs directly from plant genomic DNA. Our framework consists of two subtasks known as splicing site detection (SSD) and splicing site pairing (SSP). The model's effectiveness has been validated on gene data of 10 plant species. Trained on known circRNA instances, it is also capable of discovering previously unannotated circRNAs. In addition, we performed interpretability analyses on the trained model to investigate the sequence patterns contributing to its predictions. Our framework provides a fast and accurate computational method and tool for large-scale circRNA discovery in plants, laying a foundation for future research in plant functional genomics and non-coding RNA annotation.
☆ Recursive Reward Aggregation
In reinforcement learning (RL), aligning agent behavior with specific objectives typically requires careful design of the reward function, which can be challenging when the desired objectives are complex. In this work, we propose an alternative approach for flexible behavior alignment that eliminates the need to modify the reward function by selecting appropriate reward aggregation functions. By introducing an algebraic perspective on Markov decision processes (MDPs), we show that the Bellman equations naturally emerge from the recursive generation and aggregation of rewards, allowing for the generalization of the standard discounted sum to other recursive aggregations, such as discounted max and Sharpe ratio. Our approach applies to both deterministic and stochastic settings and integrates seamlessly with value-based and actor-critic algorithms. Experimental results demonstrate that our approach effectively optimizes diverse objectives, highlighting its versatility and potential for real-world applications.
comment: Reinforcement Learning Conference 2025
☆ Data Depth as a Risk
Data depths are score functions that quantify in an unsupervised fashion how central is a point inside a distribution, with numerous applications such as anomaly detection, multivariate or functional data analysis, arising across various fields. The halfspace depth was the first depth to aim at generalising the notion of quantile beyond the univariate case. Among the existing variety of depth definitions, it remains one of the most used notions of data depth. Taking a different angle from the quantile point of view, we show that the halfspace depth can also be regarded as the minimum loss of a set of classifiers for a specific labelling of the points. By changing the loss or the set of classifiers considered, this new angle naturally leads to a family of "loss depths", extending to well-studied classifiers such as, e.g., SVM or logistic regression, among others. This framework directly inherits computational efficiency of existing machine learning algorithms as well as their fast statistical convergence rates, and opens the data depth realm to the high-dimensional setting. Furthermore, the new loss depths highlight a connection between the dataset and the right amount of complexity or simplicity of the classifiers. The simplicity of classifiers as well as the interpretation as a risk makes our new kind of data depth easy to explain, yet efficient for anomaly detection, as is shown by experiments.
☆ SFedKD: Sequential Federated Learning with Discrepancy-Aware Multi-Teacher Knowledge Distillation
Federated Learning (FL) is a distributed machine learning paradigm which coordinates multiple clients to collaboratively train a global model via a central server. Sequential Federated Learning (SFL) is a newly-emerging FL training framework where the global model is trained in a sequential manner across clients. Since SFL can provide strong convergence guarantees under data heterogeneity, it has attracted significant research attention in recent years. However, experiments show that SFL suffers from severe catastrophic forgetting in heterogeneous environments, meaning that the model tends to forget knowledge learned from previous clients. To address this issue, we propose an SFL framework with discrepancy-aware multi-teacher knowledge distillation, called SFedKD, which selects multiple models from the previous round to guide the current round of training. In SFedKD, we extend the single-teacher Decoupled Knowledge Distillation approach to our multi-teacher setting and assign distinct weights to teachers' target-class and non-target-class knowledge based on the class distributional discrepancy between teacher and student data. Through this fine-grained weighting strategy, SFedKD can enhance model training efficacy while mitigating catastrophic forgetting. Additionally, to prevent knowledge dilution, we eliminate redundant teachers for the knowledge distillation and formalize it as a variant of the maximum coverage problem. Based on the greedy strategy, we design a complementary-based teacher selection mechanism to ensure that the selected teachers achieve comprehensive knowledge space coverage while reducing communication and computational costs. Extensive experiments show that SFedKD effectively overcomes catastrophic forgetting in SFL and outperforms state-of-the-art FL methods.
☆ Efficient Deployment of Vision-Language Models on Mobile Devices: A Case Study on OnePlus 13R
Vision-Language Models (VLMs) offer promising capabilities for mobile devices, but their deployment faces significant challenges due to computational limitations and energy inefficiency, especially for real-time applications. This study provides a comprehensive survey of deployment frameworks for VLMs on mobile devices, evaluating llama.cpp, MLC-Imp, and mllm in the context of running LLaVA-1.5 7B, MobileVLM-3B, and Imp-v1.5 3B as representative workloads on a OnePlus 13R. Each deployment framework was evaluated on the OnePlus 13R while running VLMs, with measurements covering CPU, GPU, and NPU utilization, temperature, inference time, power consumption, and user experience. Benchmarking revealed critical performance bottlenecks across frameworks: CPU resources were consistently over-utilized during token generation, while GPU and NPU accelerators were largely unused. When the GPU was used, primarily for image feature extraction, it was saturated, leading to degraded device responsiveness. The study contributes framework-level benchmarks, practical profiling tools, and an in-depth analysis of hardware utilization bottlenecks, highlighting the consistent overuse of CPUs and the ineffective or unstable use of GPUs and NPUs in current deployment frameworks.
☆ SynBridge: Bridging Reaction States via Discrete Flow for Bidirectional Reaction Prediction
The essence of a chemical reaction lies in the redistribution and reorganization of electrons, which is often manifested through electron transfer or the migration of electron pairs. These changes are inherently discrete and abrupt in the physical world, such as alterations in the charge states of atoms or the formation and breaking of chemical bonds. To model the transition of states, we propose SynBridge, a bidirectional flow-based generative model to achieve multi-task reaction prediction. By leveraging a graph-to-graph transformer network architecture and discrete flow bridges between any two discrete distributions, SynBridge captures bidirectional chemical transformations between graphs of reactants and products through the bonds' and atoms' discrete states. We further demonstrate the effectiveness of our method through extensive experiments on three benchmark datasets (USPTO-50K, USPTO-MIT, Pistachio), achieving state-of-the-art performance in both forward and retrosynthesis tasks. Our ablation studies and noise scheduling analysis reveal the benefits of structured diffusion over discrete spaces for reaction prediction.
comment: 22pages, 2 figures
☆ Evaluating SAE interpretability without explanations
Sparse autoencoders (SAEs) and transcoders have become important tools for machine learning interpretability. However, measuring how interpretable they are remains challenging, with weak consensus about which benchmarks to use. Most evaluation procedures start by producing a single-sentence explanation for each latent. These explanations are then evaluated based on how well they enable an LLM to predict the activation of a latent in new contexts. This method makes it difficult to disentangle the explanation generation and evaluation process from the actual interpretability of the latents discovered. In this work, we adapt existing methods to assess the interpretability of sparse coders, with the advantage that they do not require generating natural language explanations as an intermediate step. This enables a more direct and potentially standardized assessment of interpretability. Furthermore, we compare the scores produced by our interpretability metrics with human evaluations across similar tasks and varying setups, offering suggestions for the community on improving the evaluation of these techniques.
Pre-Training LLMs on a budget: A comparison of three optimizers
Optimizers play a decisive role in reducing pre-training times for LLMs and achieving better-performing models. In this study, we compare three major variants: the de-facto standard AdamW, the simpler Lion, developed through an evolutionary search, and the second-order optimizer Sophia. For better generalization, we train with two different base architectures and use a single- and a multiple-epoch approach while keeping the number of tokens constant. Using the Maximal Update Parametrization and smaller proxy models, we tune relevant hyperparameters separately for each combination of base architecture and optimizer. We found that while the results from all three optimizers were in approximately the same range, Sophia exhibited the lowest training and validation loss, Lion was fastest in terms of training GPU hours but AdamW led to the best downstream evaluation results.
☆ Ranked Set Sampling-Based Multilayer Perceptron: Improving Generalization via Variance-Based Bounds
Multilayer perceptron (MLP), one of the most fundamental neural networks, is extensively utilized for classification and regression tasks. In this paper, we establish a new generalization error bound, which reveals how the variance of empirical loss influences the generalization ability of the learning model. Inspired by this learning bound, we advocate to reduce the variance of empirical loss to enhance the ability of MLP. As is well-known, bagging is a popular ensemble method to realize variance reduction. However, bagging produces the base training data sets by the Simple Random Sampling (SRS) method, which exhibits a high degree of randomness. To handle this issue, we introduce an ordered structure in the training data set by Rank Set Sampling (RSS) to further reduce the variance of loss and develop a RSS-MLP method. Theoretical results show that the variance of empirical exponential loss and the logistic loss estimated by RSS are smaller than those estimated by SRS, respectively. To validate the performance of RSS-MLP, we conduct comparison experiments on twelve benchmark data sets in terms of the two convex loss functions under two fusion methods. Extensive experimental results and analysis illustrate the effectiveness and rationality of the propose method.
☆ Space filling positionality and the Spiroformer
Transformers excel when dealing with sequential data. Generalizing transformer models to geometric domains, such as manifolds, we encounter the problem of not having a well-defined global order. We propose a solution with attention heads following a space-filling curve. As a first experimental example, we present the Spiroformer, a transformer that follows a polar spiral on the $2$-sphere.
comment: 9 pages, 5 figures. To appear in Geometric Science of Information 2025
☆ Why this and not that? A Logic-based Framework for Contrastive Explanations
We define several canonical problems related to contrastive explanations, each answering a question of the form ''Why P but not Q?''. The problems compute causes for both P and Q, explicitly comparing their differences. We investigate the basic properties of our definitions in the setting of propositional logic. We show, inter alia, that our framework captures a cardinality-minimal version of existing contrastive explanations in the literature. Furthermore, we provide an extensive analysis of the computational complexities of the problems. We also implement the problems for CNF-formulas using answer set programming and present several examples demonstrating how they work in practice.
comment: 20 pages, accepted to JELIA 2025
☆ KGRAG-Ex: Explainable Retrieval-Augmented Generation with Knowledge Graph-based Perturbations
Retrieval-Augmented Generation (RAG) enhances language models by grounding responses in external information, yet explainability remains a critical challenge, particularly when retrieval relies on unstructured text. Knowledge graphs (KGs) offer a solution by introducing structured, semantically rich representations of entities and their relationships, enabling transparent retrieval paths and interpretable reasoning. In this work, we present KGRAG-Ex, a RAG system that improves both factual grounding and explainability by leveraging a domain-specific KG constructed via prompt-based information extraction. Given a user query, KGRAG-Ex identifies relevant entities and semantic paths in the graph, which are then transformed into pseudo-paragraphs: natural language representations of graph substructures that guide corpus retrieval. To improve interpretability and support reasoning transparency, we incorporate perturbation-based explanation methods that assess the influence of specific KG-derived components on the generated answers. We conduct a series of experiments to analyze the sensitivity of the system to different perturbation methods, the relationship between graph component importance and their structural positions, the influence of semantic node types, and how graph metrics correspond to the influence of components within the explanations process.
☆ Optimal and Practical Batched Linear Bandit Algorithm ICML 2025
We study the linear bandit problem under limited adaptivity, known as the batched linear bandit. While existing approaches can achieve near-optimal regret in theory, they are often computationally prohibitive or underperform in practice. We propose \texttt{BLAE}, a novel batched algorithm that integrates arm elimination with regularized G-optimal design, achieving the minimax optimal regret (up to logarithmic factors in $T$) in both large-$K$ and small-$K$ regimes for the first time, while using only $O(\log\log T)$ batches. Our analysis introduces new techniques for batch-wise optimal design and refined concentration bounds. Crucially, \texttt{BLAE} demonstrates low computational overhead and strong empirical performance, outperforming state-of-the-art methods in extensive numerical evaluations. Thus, \texttt{BLAE} is the first algorithm to combine provable minimax-optimality in all regimes and practical superiority in batched linear bandits.
comment: Accepted at ICML 2025
☆ RTNinja: a generalized machine learning framework for analyzing random telegraph noise signals in nanoelectronic devices
Random telegraph noise is a prevalent variability phenomenon in nanoelectronic devices, arising from stochastic carrier exchange at defect sites and critically impacting device reliability and performance. Conventional analysis techniques often rely on restrictive assumptions or manual interventions, limiting their applicability to complex, noisy datasets. Here, we introduce RTNinja, a generalized, fully automated machine learning framework for the unsupervised analysis of random telegraph noise signals. RTNinja deconvolves complex signals to identify the number and characteristics of hidden individual sources, without requiring prior knowledge of the system. The framework comprises two modular components: LevelsExtractor, which uses Bayesian inference and model selection to denoise and discretize the signal; and SourcesMapper, which infers source configurations through probabilistic clustering and optimization. To evaluate performance, we developed a Monte Carlo simulator that generates labeled datasets spanning broad signal-to-noise ratios and source complexities; across 7000 such datasets, RTNinja consistently demonstrated high-fidelity signal reconstruction and accurate extraction of source amplitudes and activity patterns. Our results demonstrate that RTNinja offers a robust, scalable, and device-agnostic tool for random telegraph noise characterization, enabling large-scale statistical benchmarking, reliability-centric technology qualification, predictive failure modeling, and device physics exploration in next-generation nanoelectronics.
☆ Towards AI-Native RAN: An Operator's Perspective of 6G Day 1 Standardization
Artificial Intelligence/Machine Learning (AI/ML) has become the most certain and prominent feature of 6G mobile networks. Unlike 5G, where AI/ML was not natively integrated but rather an add-on feature over existing architecture, 6G shall incorporate AI from the onset to address its complexity and support ubiquitous AI applications. Based on our extensive mobile network operation and standardization experience from 2G to 5G, this paper explores the design and standardization principles of AI-Native radio access networks (RAN) for 6G, with a particular focus on its critical Day 1 architecture, functionalities and capabilities. We investigate the framework of AI-Native RAN and present its three essential capabilities to shed some light on the standardization direction; namely, AI-driven RAN processing/optimization/automation, reliable AI lifecycle management (LCM), and AI-as-a-Service (AIaaS) provisioning. The standardization of AI-Native RAN, in particular the Day 1 features, including an AI-Native 6G RAN architecture, were proposed. For validation, a large-scale field trial with over 5000 5G-A base stations have been built and delivered significant improvements in average air interface latency, root cause identification, and network energy consumption with the proposed architecture and the supporting AI functions. This paper aims to provide a Day 1 framework for 6G AI-Native RAN standardization design, balancing technical innovation with practical deployment.
☆ SPINT: Spatial Permutation-Invariant Neural Transformer for Consistent Intracortical Motor Decoding
Intracortical Brain-Computer Interfaces (iBCI) aim to decode behavior from neural population activity, enabling individuals with motor impairments to regain motor functions and communication abilities. A key challenge in long-term iBCI is the nonstationarity of neural recordings, where the composition and tuning profiles of the recorded populations are unstable across recording sessions. Existing methods attempt to address this issue by explicit alignment techniques; however, they rely on fixed neural identities and require test-time labels or parameter updates, limiting their generalization across sessions and imposing additional computational burden during deployment. In this work, we introduce SPINT - a Spatial Permutation-Invariant Neural Transformer framework for behavioral decoding that operates directly on unordered sets of neural units. Central to our approach is a novel context-dependent positional embedding scheme that dynamically infers unit-specific identities, enabling flexible generalization across recording sessions. SPINT supports inference on variable-size populations and allows few-shot, gradient-free adaptation using a small amount of unlabeled data from the test session. To further promote model robustness to population variability, we introduce dynamic channel dropout, a regularization method for iBCI that simulates shifts in population composition during training. We evaluate SPINT on three multi-session datasets from the FALCON Benchmark, covering continuous motor decoding tasks in human and non-human primates. SPINT demonstrates robust cross-session generalization, outperforming existing zero-shot and few-shot unsupervised baselines while eliminating the need for test-time alignment and fine-tuning. Our work contributes an initial step toward a robust and scalable neural decoding framework for long-term iBCI applications.
☆ Inference-Time Scaling of Diffusion Language Models with Particle Gibbs Sampling
Discrete diffusion models have emerged as a powerful paradigm for language modeling, rivaling auto-regressive models by training-time scaling. However, inference-time scaling in discrete diffusion models remains relatively under-explored. In this work, we study sampling-based approaches for achieving high-quality text generation from discrete diffusion models in reward-guided settings. We introduce a novel inference-time scaling approach based on particle Gibbs sampling for discrete diffusion models. The particle Gibbs sampling algorithm iteratively refines full diffusion trajectories using conditional Sequential Monte Carlo as its transition mechanism. This process ensures that the updated samples progressively improve and move closer to the reward-weighted target distribution. Unlike existing inference-time scaling methods, which are often limited to single diffusion trajectories, our approach leverages iterative refinement across multiple trajectories. Within this framework, we further analyze the trade-offs between four key axes for inference-time scaling under fixed compute budgets: particle Gibbs iterations, particle count, denoising steps, and reward estimation cost. Empirically, our method consistently outperforms prior inference-time strategies on reward-guided text generation tasks, achieving significant improvement in accuracy under varying compute budgets.
☆ Online Pre-Training for Offline-to-Online Reinforcement Learning ICML 2025
Offline-to-online reinforcement learning (RL) aims to integrate the complementary strengths of offline and online RL by pre-training an agent offline and subsequently fine-tuning it through online interactions. However, recent studies reveal that offline pre-trained agents often underperform during online fine-tuning due to inaccurate value estimation caused by distribution shift, with random initialization proving more effective in certain cases. In this work, we propose a novel method, Online Pre-Training for Offline-to-Online RL (OPT), explicitly designed to address the issue of inaccurate value estimation in offline pre-trained agents. OPT introduces a new learning phase, Online Pre-Training, which allows the training of a new value function tailored specifically for effective online fine-tuning. Implementation of OPT on TD3 and SPOT demonstrates an average 30% improvement in performance across a wide range of D4RL environments, including MuJoCo, Antmaze, and Adroit.
comment: ICML 2025 camera-ready
☆ Two-cluster test
Cluster analysis is a fundamental research issue in statistics and machine learning. In many modern clustering methods, we need to determine whether two subsets of samples come from the same cluster. Since these subsets are usually generated by certain clustering procedures, the deployment of classic two-sample tests in this context would yield extremely smaller p-values, leading to inflated Type-I error rate. To overcome this bias, we formally introduce the two-cluster test issue and argue that it is a totally different significance testing issue from conventional two-sample test. Meanwhile, we present a new method based on the boundary points between two subsets to derive an analytical p-value for the purpose of significance quantification. Experiments on both synthetic and real data sets show that the proposed test is able to significantly reduce the Type-I error rate, in comparison with several classic two-sample testing methods. More importantly, the practical usage of such two-cluster test is further verified through its applications in tree-based interpretable clustering and significance-based hierarchical clustering.
☆ Advances in Machine Learning: Where Can Quantum Techniques Help?
Quantum Machine Learning (QML) represents a promising frontier at the intersection of quantum computing and artificial intelligence, aiming to leverage quantum computational advantages to enhance data-driven tasks. This review explores the potential of QML to address the computational bottlenecks of classical machine learning, particularly in processing complex datasets. We introduce the theoretical foundations of QML, including quantum data encoding, quantum learning theory and optimization techniques, while categorizing QML approaches based on data type and computational architecture. It is well-established that quantum computational advantages are problem-dependent, and so potentially useful directions for QML need to be systematically identified. Key developments, such as Quantum Principal Component Analysis, quantum-enhanced sensing and applications in material science, are critically evaluated for their theoretical speed-ups and practical limitations. The challenges posed by Noisy Intermediate-Scale Quantum (NISQ) devices, including hardware noise, scalability constraints and data encoding overheads, are discussed in detail. We also outline future directions, emphasizing the need for quantum-native algorithms, improved error correction, and realistic benchmarks to bridge the gap between theoretical promise and practical deployment. This comprehensive analysis underscores that while QML has significant potential for specific applications such as quantum chemistry and sensing, its broader utility in real-world scenarios remains contingent on overcoming technological and methodological hurdles.
comment: 28 pages, 1 figure
☆ Prediction of Lane Change Intentions of Human Drivers using an LSTM, a CNN and a Transformer
Lane changes of preceding vehicles have a great impact on the motion planning of automated vehicles especially in complex traffic situations. Predicting them would benefit the public in terms of safety and efficiency. While many research efforts have been made in this direction, few concentrated on predicting maneuvers within a set time interval compared to predicting at a set prediction time. In addition, there exist a lack of comparisons between different architectures to try to determine the best performing one and to assess how to correctly choose the input for such models. In this paper the structure of an LSTM, a CNN and a Transformer network are described and implemented to predict the intention of human drivers to perform a lane change. We show how the data was prepared starting from a publicly available dataset (highD), which features were used, how the networks were designed and finally we compare the results of the three networks with different configurations of input data. We found that transformer networks performed better than the other networks and was less affected by overfitting. The accuracy of the method spanned from $82.79\%$ to $96.73\%$ for different input configurations and showed overall good performances considering also precision and recall.
comment: 14 pages, 18 figures
☆ Leveraging Machine Learning and Enhanced Parallelism Detection for BPMN Model Generation from Text
Efficient planning, resource management, and consistent operations often rely on converting textual process documents into formal Business Process Model and Notation (BPMN) models. However, this conversion process remains time-intensive and costly. Existing approaches, whether rule-based or machine-learning-based, still struggle with writing styles and often fail to identify parallel structures in process descriptions. This paper introduces an automated pipeline for extracting BPMN models from text, leveraging the use of machine learning and large language models. A key contribution of this work is the introduction of a newly annotated dataset, which significantly enhances the training process. Specifically, we augment the PET dataset with 15 newly annotated documents containing 32 parallel gateways for model training, a critical feature often overlooked in existing datasets. This addition enables models to better capture parallel structures, a common but complex aspect of process descriptions. The proposed approach demonstrates adequate performance in terms of reconstruction accuracy, offering a promising foundation for organizations to accelerate BPMN model creation.
☆ scE$^2$TM: Toward Interpretable Single-Cell Embedding via Topic Modeling
Recent advances in sequencing technologies have enabled researchers to explore cellular heterogeneity at single-cell resolution. Meanwhile, interpretability has gained prominence parallel to the rapid increase in the complexity and performance of deep learning models. In recent years, topic models have been widely used for interpretable single-cell embedding learning and clustering analysis, which we refer to as single-cell embedded topic models. However, previous studies evaluated the interpretability of the models mainly through qualitative analysis, and these single-cell embedded topic models suffer from the potential problem of interpretation collapse. Furthermore, their neglect of external biological knowledge constrains analytical performance. Here, we present scE2TM, an external knowledge-guided single-cell embedded topic model that provides a high-quality cell embedding and strong interpretation, contributing to comprehensive scRNA-seq data analysis. Our comprehensive evaluation across 20 scRNA-seq datasets demonstrates that scE2TM achieves significant clustering performance gains compared to 7 state-of-the-art methods. In addition, we propose a new interpretability evaluation benchmark that introduces 10 metrics to quantitatively assess the interpretability of single-cell embedded topic models. The results show that the interpretation provided by scE2TM performs encouragingly in terms of diversity and consistency with the underlying biological signals, contributing to a better revealing of the underlying biological mechanisms.
☆ Audio Inpanting using Discrete Diffusion Model
Audio inpainting refers to the task of reconstructing missing segments in corrupted audio recordings. While prior approaches-including waveform and spectrogram-based diffusion models-have shown promising results for short gaps, they often degrade in quality when gaps exceed 100 milliseconds (ms). In this work, we introduce a novel inpainting method based on discrete diffusion modeling, which operates over tokenized audio representations produced by a pre-trained audio tokenizer. Our approach models the generative process directly in the discrete latent space, enabling stable and semantically coherent reconstruction of missing audio. We evaluate the method on the MusicNet dataset using both objective and perceptual metrics across gap durations up to 300 ms. We further evaluated our approach on the MTG dataset, extending the gap duration to 500 ms. Experimental results demonstrate that our method achieves competitive or superior performance compared to existing baselines, particularly for longer gaps, offering a robust solution for restoring degraded musical recordings. Audio examples of our proposed method can be found at https://iftach21.github.io/
☆ Interpretability-Aware Pruning for Efficient Medical Image Analysis
Deep learning has driven significant advances in medical image analysis, yet its adoption in clinical practice remains constrained by the large size and lack of transparency in modern models. Advances in interpretability techniques such as DL-Backtrace, Layer-wise Relevance Propagation, and Integrated Gradients make it possible to assess the contribution of individual components within neural networks trained on medical imaging tasks. In this work, we introduce an interpretability-guided pruning framework that reduces model complexity while preserving both predictive performance and transparency. By selectively retaining only the most relevant parts of each layer, our method enables targeted compression that maintains clinically meaningful representations. Experiments across multiple medical image classification benchmarks demonstrate that this approach achieves high compression rates with minimal loss in accuracy, paving the way for lightweight, interpretable models suited for real-world deployment in healthcare settings.
comment: Pre-Print
☆ Towards Efficient Quantity Retrieval from Text:an Approach via Description Parsing and Weak Supervision
Quantitative facts are continually generated by companies and governments, supporting data-driven decision-making. While common facts are structured, many long-tail quantitative facts remain buried in unstructured documents, making them difficult to access. We propose the task of Quantity Retrieval: given a description of a quantitative fact, the system returns the relevant value and supporting evidence. Understanding quantity semantics in context is essential for this task. We introduce a framework based on description parsing that converts text into structured (description, quantity) pairs for effective retrieval. To improve learning, we construct a large paraphrase dataset using weak supervision based on quantity co-occurrence. We evaluate our approach on a large corpus of financial annual reports and a newly annotated quantity description dataset. Our method significantly improves top-1 retrieval accuracy from 30.98 percent to 64.66 percent.
comment: Extended version of the paper accepted in DEXA 2025
☆ A Comprehensively Adaptive Architectural Optimization-Ingrained Quantum Neural Network Model for Cloud Workloads Prediction
Accurate workload prediction and advanced resource reservation are indispensably crucial for managing dynamic cloud services. Traditional neural networks and deep learning models frequently encounter challenges with diverse, high-dimensional workloads, especially during sudden resource demand changes, leading to inefficiencies. This issue arises from their limited optimization during training, relying only on parametric (inter-connection weights) adjustments using conventional algorithms. To address this issue, this work proposes a novel Comprehensively Adaptive Architectural Optimization-based Variable Quantum Neural Network (CA-QNN), which combines the efficiency of quantum computing with complete structural and qubit vector parametric learning. The model converts workload data into qubits, processed through qubit neurons with Controlled NOT-gated activation functions for intuitive pattern recognition. In addition, a comprehensive architecture optimization algorithm for networks is introduced to facilitate the learning and propagation of the structure and parametric values in variable-sized QNNs. This algorithm incorporates quantum adaptive modulation and size-adaptive recombination during training process. The performance of CA-QNN model is thoroughly investigated against seven state-of-the-art methods across four benchmark datasets of heterogeneous cloud workloads. The proposed model demonstrates superior prediction accuracy, reducing prediction errors by up to 93.40% and 91.27% compared to existing deep learning and QNN-based approaches.
☆ CAS Condensed and Accelerated Silhouette: An Efficient Method for Determining the Optimal K in K-Means Clustering
Clustering is a critical component of decision-making in todays data-driven environments. It has been widely used in a variety of fields such as bioinformatics, social network analysis, and image processing. However, clustering accuracy remains a major challenge in large datasets. This paper presents a comprehensive overview of strategies for selecting the optimal value of k in clustering, with a focus on achieving a balance between clustering precision and computational efficiency in complex data environments. In addition, this paper introduces improvements to clustering techniques for text and image data to provide insights into better computational performance and cluster validity. The proposed approach is based on the Condensed Silhouette method, along with statistical methods such as Local Structures, Gap Statistics, Class Consistency Ratio, and a Cluster Overlap Index CCR and COIbased algorithm to calculate the best value of k for K-Means clustering. The results of comparative experiments show that the proposed approach achieves up to 99 percent faster execution times on high-dimensional datasets while retaining both precision and scalability, making it highly suitable for real time clustering needs or scenarios demanding efficient clustering with minimal resource utilization.
☆ M2-Reasoning: Empowering MLLMs with Unified General and Spatial Reasoning
Recent advancements in Multimodal Large Language Models (MLLMs), particularly through Reinforcement Learning with Verifiable Rewards (RLVR), have significantly enhanced their reasoning abilities. However, a critical gap persists: these models struggle with dynamic spatial interactions, a capability essential for real-world applications. To bridge this gap, we introduce M2-Reasoning-7B, a model designed to excel in both general and spatial reasoning. Our approach integrates two key innovations: (1) a novel data pipeline that generates 294.2K high-quality data samples (168K for cold-start fine-tuning and 126.2K for RLVR), which feature logically coherent reasoning trajectories and have undergone comprehensive assessment; and (2) a dynamic multi-task training strategy with step-wise optimization to mitigate conflicts between data, and task-specific rewards for delivering tailored incentive signals. This combination of curated data and advanced training allows M2-Reasoning-7B to set a new state-of-the-art (SOTA) across 8 benchmarks, showcasing superior performance in both general and spatial reasoning domains.
comment: 31pages, 14 figures
☆ Lightweight Safety Guardrails via Synthetic Data and RL-guided Adversarial Training
We introduce a lightweight yet highly effective safety guardrail framework for language models, demonstrating that small-scale language models can achieve, and even surpass, the performance of larger counterparts in content moderation tasks. This is accomplished through high-fidelity synthetic data generation and adversarial training. The synthetic data generation process begins with human-curated seed data, which undergoes query augmentation and paraphrasing to create diverse and contextually rich examples. This augmented data is then subjected to multiple rounds of curation, ensuring high fidelity and relevance. Inspired by recent advances in the Generative Adversarial Network (GAN) architecture, our adversarial training employs reinforcement learning to guide a generator that produces challenging synthetic examples. These examples are used to fine-tune the safety classifier, enhancing its ability to detect and mitigate harmful content. Additionally, we incorporate strategies from recent research on efficient LLM training, leveraging the capabilities of smaller models to improve the performance of larger generative models. With iterative adversarial training and the generation of diverse, high-quality synthetic data, our framework enables small language models (SLMs) to serve as robust safety guardrails. This approach not only reduces computational overhead but also enhances resilience against adversarial attacks, offering a scalable and efficient solution for content moderation in AI systems.
☆ MIRRAMS: Towards Training Models Robust to Missingness Distribution Shifts
In real-world data analysis, missingness distributional shifts between training and test input datasets frequently occur, posing a significant challenge to achieving robust prediction performance. In this study, we propose a novel deep learning framework designed to address such shifts in missingness distributions. We begin by introducing a set of mutual information-based conditions, called MI robustness conditions, which guide a prediction model to extract label-relevant information while remaining invariant to diverse missingness patterns, thereby enhancing robustness to unseen missingness scenarios at test-time. To make these conditions practical, we propose simple yet effective techniques to derive loss terms corresponding to each and formulate a final objective function, termed MIRRAMS(Mutual Information Regularization for Robustness Against Missingness Shifts). As a by-product, our analysis provides a theoretical interpretation of the principles underlying consistency regularization-based semi-supervised learning methods, such as FixMatch. Extensive experiments across various benchmark datasets show that MIRRAMS consistently outperforms existing baselines and maintains stable performance across diverse missingness scenarios. Moreover, our approach achieves state-of-the-art performance even without missing data and can be naturally extended to address semi-supervised learning tasks, highlighting MIRRAMS as a powerful, off-the-shelf framework for general-purpose learning.
☆ Data-Driven Dimensional Synthesis of Diverse Planar Four-bar Function Generation Mechanisms via Direct Parameterization
Dimensional synthesis of planar four-bar mechanisms is a challenging inverse problem in kinematics, requiring the determination of mechanism dimensions from desired motion specifications. We propose a data-driven framework that bypasses traditional equation-solving and optimization by leveraging supervised learning. Our method combines a synthetic dataset, an LSTM-based neural network for handling sequential precision points, and a Mixture of Experts (MoE) architecture tailored to different linkage types. Each expert model is trained on type-specific data and guided by a type-specifying layer, enabling both single-type and multi-type synthesis. A novel simulation metric evaluates prediction quality by comparing desired and generated motions. Experiments show our approach produces accurate, defect-free linkages across various configurations. This enables intuitive and efficient mechanism design, even for non-expert users, and opens new possibilities for scalable and flexible synthesis in kinematic design.
☆ A Practical Two-Stage Recipe for Mathematical LLMs: Maximizing Accuracy with SFT and Efficiency with Reinforcement Learning ICML 2025
Enhancing the mathematical reasoning of Large Language Models (LLMs) is a pivotal challenge in advancing AI capabilities. While Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) are the dominant training paradigms, a systematic methodology for combining them to maximize both accuracy and efficiency remains largely unexplored. This paper introduces a practical and effective training recipe that strategically integrates extended SFT with RL from online inference (GRPO). We posit that these methods play complementary, not competing, roles: a prolonged SFT phase first pushes the model's accuracy to its limits, after which a GRPO phase dramatically improves token efficiency while preserving this peak performance. Our experiments reveal that extending SFT for as many as 10 epochs is crucial for performance breakthroughs, and that the primary role of GRPO in this framework is to optimize solution length. The efficacy of our recipe is rigorously validated through top-tier performance on challenging benchmarks, including a high rank among over 2,200 teams in the strictly leak-free AI Mathematical Olympiad (AIMO). This work provides the community with a battle-tested blueprint for developing state-of-the-art mathematical reasoners that are both exceptionally accurate and practically efficient. To ensure full reproducibility and empower future research, we will open-source our entire framework, including all code, model checkpoints, and training configurations at https://github.com/analokmaus/kaggle-aimo2-fast-math-r1.
comment: Presented at ICML 2025 Workshop on The second AI for MATH
☆ Admissibility of Stein Shrinkage for Batch Normalization in the Presence of Adversarial Attacks
Batch normalization (BN) is a ubiquitous operation in deep neural networks used primarily to achieve stability and regularization during network training. BN involves feature map centering and scaling using sample means and variances, respectively. Since these statistics are being estimated across the feature maps within a batch, this problem is ideally suited for the application of Stein's shrinkage estimation, which leads to a better, in the mean-squared-error sense, estimate of the mean and variance of the batch. In this paper, we prove that the Stein shrinkage estimator for the mean and variance dominates over the sample mean and variance estimators in the presence of adversarial attacks when modeling these attacks using sub-Gaussian distributions. This facilitates and justifies the application of Stein shrinkage to estimate the mean and variance parameters in BN and use it in image classification (segmentation) tasks with and without adversarial attacks. We present SOTA performance results using this Stein corrected batch norm in a standard ResNet architecture applied to the task of image classification using CIFAR-10 data, 3D CNN on PPMI (neuroimaging) data and image segmentation using HRNet on Cityscape data with and without adversarial attacks.
☆ Quantum-Accelerated Neural Imputation with Large Language Models (LLMs)
Missing data presents a critical challenge in real-world datasets, significantly degrading the performance of machine learning models. While Large Language Models (LLMs) have recently demonstrated remarkable capabilities in tabular data imputation, exemplified by frameworks like UnIMP, their reliance on classical embedding methods often limits their ability to capture complex, non-linear correlations, particularly in mixed-type data scenarios encompassing numerical, categorical, and textual features. This paper introduces Quantum-UnIMP, a novel framework that integrates shallow quantum circuits into an LLM-based imputation architecture. Our core innovation lies in replacing conventional classical input embeddings with quantum feature maps generated by an Instantaneous Quantum Polynomial (IQP) circuit. This approach enables the model to leverage quantum phenomena such as superposition and entanglement, thereby learning richer, more expressive representations of data and enhancing the recovery of intricate missingness patterns. Our experiments on benchmark mixed-type datasets demonstrate that Quantum-UnIMP reduces imputation error by up to 15.2% for numerical features (RMSE) and improves classification accuracy by 8.7% for categorical features (F1-Score) compared to state-of-the-art classical and LLM-based methods. These compelling results underscore the profound potential of quantum-enhanced representations for complex data imputation tasks, even with near-term quantum hardware.
☆ Raptor: Scalable Train-Free Embeddings for 3D Medical Volumes Leveraging Pretrained 2D Foundation Models ICML 2025
Current challenges in developing foundational models for volumetric imaging data, such as magnetic resonance imaging (MRI), stem from the computational complexity of training state-of-the-art architectures in high dimensions and curating sufficiently large datasets of volumes. To address these challenges, we introduce Raptor (Random Planar Tensor Reduction), a train-free method for generating semantically rich embeddings for volumetric data. Raptor leverages a frozen 2D foundation model, pretrained on natural images, to extract visual tokens from individual cross-sections of medical volumes. These tokens are then spatially compressed using random projections, significantly reducing computational complexity while retaining semantic information. Extensive experiments on ten diverse medical volume tasks verify the superior performance of Raptor over state-of-the-art methods, including those pretrained exclusively on medical volumes (+3% SuPreM, +6% MISFM, +10% Merlin, +13% VoCo, and +14% SLIViT), while entirely bypassing the need for costly training. Our results highlight the effectiveness and versatility of Raptor as a foundation for advancing deep learning-based methods for medical volumes.
comment: 21 pages, 10 figures, accepted to ICML 2025. The first two authors contributed equally
☆ Transfer Learning and Mixup for Fine-Grained Few-Shot Fungi Classification
Accurate identification of fungi species presents a unique challenge in computer vision due to fine-grained inter-species variation and high intra-species variation. This paper presents our approach for the FungiCLEF 2025 competition, which focuses on few-shot fine-grained visual categorization (FGVC) using the FungiTastic Few-Shot dataset. Our team (DS@GT) experimented with multiple vision transformer models, data augmentation, weighted sampling, and incorporating textual information. We also explored generative AI models for zero-shot classification using structured prompting but found them to significantly underperform relative to vision-based models. Our final model outperformed both competition baselines and highlighted the effectiveness of domain specific pretraining and balanced sampling strategies. Our approach ranked 35/74 on the private test set in post-completion evaluation, this suggests additional work can be done on metadata selection and domain-adapted multi-modal learning. Our code is available at https://github.com/dsgt-arc/fungiclef-2025.
☆ CoreSPECT: Enhancing Clustering Algorithms via an Interplay of Density and Geometry
Density and geometry have long served as two of the fundamental guiding principles in clustering algorithm design, with algorithm usually focusing either on the density structure of the data (e.g., HDBSCAN and Density Peak Clustering) or the complexity of underlying geometry (e.g., manifold clustering algorithms). In this paper, we identify and formalize a recurring but often overlooked interaction between distribution and geometry and leverage this insight to design our clustering enhancement framework CoreSPECT (Core Space Projection-based Enhancement of Clustering Techniques). Our framework boosts the performance of simple algorithms like K-Means and GMM by applying them to strategically selected regions, then extending the partial partition to a complete partition for the dataset using a novel neighborhood graph based multi-layer propagation procedure. We apply our framework on 15 datasets from three different domains and obtain consistent and substantial gain in clustering accuracy for both K-Means and GMM. On average, our framework improves the ARI of K-Means by 40% and of GMM by 14%, often surpassing the performance of both manifold-based and recent density-based clustering algorithms. We further support our framework with initial theoretical guarantees, ablation to demonstrate the usefulness of the individual steps and with evidence of robustness to noise.
☆ Exploring Gender Differences in Chronic Pain Discussions on Reddit
Pain is an inherent part of human existence, manifesting as both physical and emotional experiences, and can be categorized as either acute or chronic. Over the years, extensive research has been conducted to understand the causes of pain and explore potential treatments, with contributions from various scientific disciplines. However, earlier studies often overlooked the role of gender in pain experiences. In this study, we utilized Natural Language Processing (NLP) to analyze and gain deeper insights into individuals' pain experiences, with a particular focus on gender differences. We successfully classified posts into male and female corpora using the Hidden Attribute Model-Convolutional Neural Network (HAM-CNN), achieving an F1 score of 0.86 by aggregating posts based on usernames. Our analysis revealed linguistic differences between genders, with female posts tending to be more emotionally focused. Additionally, the study highlighted that conditions such as migraine and sinusitis are more prevalent among females and explored how pain medication affects individuals differently based on gender.
comment: This is an extended version of the short paper accepted at ASONAM 2025
☆ Data Generation without Function Estimation
Estimating the score function (or other population-density-dependent functions) is a fundamental component of most generative models. However, such function estimation is computationally and statistically challenging. Can we avoid function estimation for data generation? We propose an estimation-free generative method: A set of points whose locations are deterministically updated with (inverse) gradient descent can transport a uniform distribution to arbitrary data distribution, in the mean field regime, without function estimation, training neural networks, and even noise injection. The proposed method is built upon recent advances in the physics of interacting particles. We show, both theoretically and experimentally, that these advances can be leveraged to develop novel generative methods.
Self-Supervised Learning-Based Multimodal Prediction on Prosocial Behavior Intentions ICASSP 2025
Human state detection and behavior prediction have seen significant advancements with the rise of machine learning and multimodal sensing technologies. However, predicting prosocial behavior intentions in mobility scenarios, such as helping others on the road, is an underexplored area. Current research faces a major limitation. There are no large, labeled datasets available for prosocial behavior, and small-scale datasets make it difficult to train deep-learning models effectively. To overcome this, we propose a self-supervised learning approach that harnesses multi-modal data from existing physiological and behavioral datasets. By pre-training our model on diverse tasks and fine-tuning it with a smaller, manually labeled prosocial behavior dataset, we significantly enhance its performance. This method addresses the data scarcity issue, providing a more effective benchmark for prosocial behavior prediction, and offering valuable insights for improving intelligent vehicle systems and human-machine interaction.
comment: 5 pages, 4 figures, published at ICASSP 2025
☆ InsightBuild: LLM-Powered Causal Reasoning in Smart Building Systems
Smart buildings generate vast streams of sensor and control data, but facility managers often lack clear explanations for anomalous energy usage. We propose InsightBuild, a two-stage framework that integrates causality analysis with a fine-tuned large language model (LLM) to provide human-readable, causal explanations of energy consumption patterns. First, a lightweight causal inference module applies Granger causality tests and structural causal discovery on building telemetry (e.g., temperature, HVAC settings, occupancy) drawn from Google Smart Buildings and Berkeley Office datasets. Next, an LLM, fine-tuned on aligned pairs of sensor-level causes and textual explanations, receives as input the detected causal relations and generates concise, actionable explanations. We evaluate InsightBuild on two real-world datasets (Google: 2017-2022; Berkeley: 2018-2020), using expert-annotated ground-truth causes for a held-out set of anomalies. Our results demonstrate that combining explicit causal discovery with LLM-based natural language generation yields clear, precise explanations that assist facility managers in diagnosing and mitigating energy inefficiencies.
♻ ☆ Drowning in Documents: Consequences of Scaling Reranker Inference SIGIR 2025
Rerankers, typically cross-encoders, are computationally intensive but are frequently used because they are widely assumed to outperform cheaper initial IR systems. We challenge this assumption by measuring reranker performance for full retrieval, not just re-scoring first-stage retrieval. To provide a more robust evaluation, we prioritize strong first-stage retrieval using modern dense embeddings and test rerankers on a variety of carefully chosen, challenging tasks, including internally curated datasets to avoid contamination, and out-of-domain ones. Our empirical results reveal a surprising trend: the best existing rerankers provide initial improvements when scoring progressively more documents, but their effectiveness gradually declines and can even degrade quality beyond a certain limit. We hope that our findings will spur future research to improve reranking.
comment: Accepted to ReNeuIR 2025 Workshop at SIGIR 2025 Conference
♻ ☆ Learning-aided Bigraph Matching Approach to Multi-Crew Restoration of Damaged Power Networks Coupled with Road Transportation Networks
The resilience of critical infrastructure networks (CINs) after disruptions, such as those caused by natural hazards, depends on both the speed of restoration and the extent to which operational functionality can be regained. Allocating resources for restoration is a combinatorial optimal planning problem that involves determining which crews will repair specific network nodes and in what order. This paper presents a novel graph-based formulation that merges two interconnected graphs, representing crew and transportation nodes and power grid nodes, into a single heterogeneous graph. To enable efficient planning, graph reinforcement learning (GRL) is integrated with bigraph matching. GRL is utilized to design the incentive function for assigning crews to repair tasks based on the graph-abstracted state of the environment, ensuring generalization across damage scenarios. Two learning techniques are employed: a graph neural network trained using Proximal Policy Optimization and another trained via Neuroevolution. The learned incentive functions inform a bipartite graph that links crews to repair tasks, enabling weighted maximum matching for crew-to-task allocations. An efficient simulation environment that pre-computes optimal node-to-node path plans is used to train the proposed restoration planning methods. An IEEE 8500-bus power distribution test network coupled with a 21 square km transportation network is used as the case study, with scenarios varying in terms of numbers of damaged nodes, depots, and crews. Results demonstrate the approach's generalizability and scalability across scenarios, with learned policies providing 3-fold better performance than random policies, while also outperforming optimization-based solutions in both computation time (by several orders of magnitude) and power restored.
comment: Accepted for presentation in proceedings of ASME IDETC 2025
♻ ☆ Exploring Efficient Quantification of Modeling Uncertainties with Differentiable Physics-Informed Machine Learning Architectures
Quantifying and propagating modeling uncertainties is crucial for reliability analysis, robust optimization, and other model-based algorithmic processes in engineering design and control. Now, physics-informed machine learning (PIML) methods have emerged in recent years as a new alternative to traditional computational modeling and surrogate modeling methods, offering a balance between computing efficiency, modeling accuracy, and interpretability. However, their ability to predict and propagate modeling uncertainties remains mostly unexplored. In this paper, a promising class of auto-differentiable hybrid PIML architectures that combine partial physics and neural networks or ANNs (for input transformation or adaptive parameter estimation) is integrated with Bayesian Neural networks (replacing the ANNs); this is done with the goal to explore whether BNNs can successfully provision uncertainty propagation capabilities in the PIML architectures as well, further supported by the auto-differentiability of these architectures. A two-stage training process is used to alleviate the challenges traditionally encountered in training probabilistic ML models. The resulting BNN-integrated PIML architecture is evaluated on an analytical benchmark problem and flight experiments data for a fixed-wing RC aircraft, with prediction performance observed to be slightly worse or at par with purely data-driven ML and original PIML models. Moreover, Monte Carlo sampling of probabilistic BNN weights was found to be most effective in propagating uncertainty in the BNN-integrated PIML architectures.
comment: Accepted for presentation in proceedings of ASME IDETC 2025
♻ ☆ Predicting Barge Presence and Quantity on Inland Waterways using Vessel Tracking Data: A Machine Learning Approach
This study presents a machine learning approach to predict the number of barges transported by vessels on inland waterways using tracking data from the Automatic Identification System (AIS). While AIS tracks the location of tug and tow vessels, it does not monitor the presence or number of barges transported by those vessels. Understanding the number and types of barges conveyed along river segments, between ports, and at ports is crucial for estimating the quantities of freight transported on the nation's waterways. This insight is also valuable for waterway management and infrastructure operations impacting areas such as targeted dredging operations, and data-driven resource allocation. Labeled sample data was generated using observations from traffic cameras located along key river segments and matched to AIS data records. A sample of 164 vessels representing up to 42 barge convoys per vessel was used for model development. The methodology involved first predicting barge presence and then predicting barge quantity. Features derived from the AIS data included speed measures, vessel characteristics, turning measures, and interaction terms. For predicting barge presence, the AdaBoost model achieved an F1 score of 0.932. For predicting barge quantity, the Random Forest combined with an AdaBoost ensemble model achieved an F1 score of 0.886. Bayesian optimization was used for hyperparameter tuning. By advancing predictive modeling for inland waterways, this study offers valuable insights for transportation planners and organizations, which require detailed knowledge of traffic volumes, including the flow of commodities, their destinations, and the tonnage moving in and out of ports.
♻ ☆ Local Flow Matching Generative Models
Flow Matching (FM) is a simulation-free method for learning a continuous and invertible flow to interpolate between two distributions, and in particular to generate data from noise. Inspired by the variational nature of the diffusion process as a gradient flow, we introduce a stepwise FM model called Local Flow Matching (LFM), which consecutively learns a sequence of FM sub-models, each matching a diffusion process up to the time of the step size in the data-to-noise direction. In each step, the two distributions to be interpolated by the sub-flow model are closer to each other than data vs. noise, and this enables the use of smaller models with faster training. This variational perspective also allows us to theoretically prove a generation guarantee of the proposed flow model in terms of the $\chi^2$-divergence between the generated and true data distributions, utilizing the contraction property of the diffusion process. In practice, the stepwise structure of LFM is natural to be distilled and different distillation techniques can be adopted to speed up generation. We empirically demonstrate improved training efficiency and competitive generative performance of LFM compared to FM on the unconditional generation of tabular data and image datasets, and also on the conditional generation of robotic manipulation policies.
♻ ☆ The Value of Prediction in Identifying the Worst-Off
Machine learning is increasingly used in government programs to identify and support the most vulnerable individuals, prioritizing assistance for those at greatest risk over optimizing aggregate outcomes. This paper examines the welfare impacts of prediction in equity-driven contexts, and how they compare to other policy levers, such as expanding bureaucratic capacity. Through mathematical models and a real-world case study on long-term unemployment amongst German residents, we develop a comprehensive understanding of the relative effectiveness of prediction in surfacing the worst-off. Our findings provide clear analytical frameworks and practical, data-driven tools that empower policymakers to make principled decisions when designing these systems.
♻ ☆ Riemannian Time Warping: Multiple Sequence Alignment in Curved Spaces
Temporal alignment of multiple signals through time warping is crucial in many fields, such as classification within speech recognition or robot motion learning. Almost all related works are limited to data in Euclidean space. Although an attempt was made in 2011 to adapt this concept to unit quaternions, a general extension to Riemannian manifolds remains absent. Given its importance for numerous applications in robotics and beyond, we introduce Riemannian Time Warping (RTW). This novel approach efficiently aligns multiple signals by considering the geometric structure of the Riemannian manifold in which the data is embedded. Extensive experiments on synthetic and real-world data, including tests with an LBR iiwa robot, demonstrate that RTW consistently outperforms state-of-the-art baselines in both averaging and classification tasks.
♻ ☆ Discovering Algorithms with Computational Language Processing
Algorithms are the engine for reproducible problem-solving. We present a framework automating algorithm discovery by conceptualizing them as sequences of operations, represented as tokens. These computational tokens are chained using a grammar, enabling the formation of increasingly sophisticated procedures. Our ensemble Monte Carlo tree search (MCTS) guided by reinforcement learning (RL) explores token chaining and drives the creation of new tokens. This methodology rediscovers, improves, and generates new algorithms that substantially outperform existing methods for strongly NP-hard combinatorial optimization problems and foundational quantum computing approaches such as Grover's and Quantum Approximate Optimization Algorithm. Operating at the computational rather than code-generation level, our framework produces algorithms that can be tailored specifically to problem instances, not merely classes.
comment: 21 pages
♻ ☆ Alternating Gradient Flows: A Theory of Feature Learning in Two-layer Neural Networks
What features neural networks learn, and how, remains an open question. In this paper, we introduce Alternating Gradient Flows (AGF), an algorithmic framework that describes the dynamics of feature learning in two-layer networks trained from small initialization. Prior works have shown that gradient flow in this regime exhibits a staircase-like loss curve, alternating between plateaus where neurons slowly align to useful directions and sharp drops where neurons rapidly grow in norm. AGF approximates this behavior as an alternating two-step process: maximizing a utility function over dormant neurons and minimizing a cost function over active ones. AGF begins with all neurons dormant. At each round, a dormant neuron activates, triggering the acquisition of a feature and a drop in the loss. AGF quantifies the order, timing, and magnitude of these drops, matching experiments across architectures. We show that AGF unifies and extends existing saddle-to-saddle analyses in fully connected linear networks and attention-only linear transformers, where the learned features are singular modes and principal components, respectively. In diagonal linear networks, we prove AGF converges to gradient flow in the limit of vanishing initialization. Applying AGF to quadratic networks trained to perform modular addition, we give the first complete characterization of the training dynamics, revealing that networks learn Fourier features in decreasing order of coefficient magnitude. Altogether, AGF offers a promising step towards understanding feature learning in neural networks.
comment: 39 pages, 7 figures
♻ ☆ On learning functions over biological sequence space: relating Gaussian process priors, regularization, and gauge fixing
Mappings from biological sequences (DNA, RNA, protein) to quantitative measures of sequence functionality play an important role in contemporary biology. We are interested in the related tasks of (i) inferring predictive sequence-to-function maps and (ii) decomposing sequence-function maps to elucidate the contributions of individual subsequences. Because each sequence-function map can be written as a weighted sum over subsequences in multiple ways, meaningfully interpreting these weights requires "gauge-fixing," i.e., defining a unique representation for each map. Recent work has established that most existing gauge-fixed representations arise as the unique solutions to $L_2$-regularized regression in an overparameterized "weight space" where the choice of regularizer defines the gauge. Here, we establish the relationship between regularized regression in overparameterized weight space and Gaussian process approaches that operate in "function space," i.e. the space of all real-valued functions on a finite set of sequences. We disentangle how weight space regularizers both impose an implicit prior on the learned function and restrict the optimal weights to a particular gauge. We also show how to construct regularizers that correspond to arbitrary explicit Gaussian process priors combined with a wide variety of gauges. Next, we derive the distribution of gauge-fixed weights implied by the Gaussian process posterior and demonstrate that even for long sequences this distribution can be efficiently computed for product-kernel priors using a kernel trick. Finally, we characterize the implicit function space priors associated with the most common weight space regularizers. Overall, our framework unifies and extends our ability to infer and interpret sequence-function relationships.
♻ ☆ Rethinking Approximate Gaussian Inference in Classification
In classification tasks, softmax functions are ubiquitously used as output activations to produce predictive probabilities. Such outputs only capture aleatoric uncertainty. To capture epistemic uncertainty, approximate Gaussian inference methods have been proposed. We develop a common formalism to describe such methods, which we view as outputting Gaussian distributions over the logit space. Predictives are then obtained as the expectations of the Gaussian distributions pushed forward through the softmax. However, such softmax Gaussian integrals cannot be solved analytically, and Monte Carlo (MC) approximations can be costly and noisy. We propose to replace the softmax activation by element-wise normCDF or sigmoid, which allows for the accurate sampling-free approximation of predictives. This also enables the approximation of the Gaussian pushforwards by Dirichlet distributions with moment matching. This approach entirely eliminates the runtime and memory overhead associated with MC sampling. We evaluate it combined with several approximate Gaussian inference methods (Laplace, HET, SNGP) on large- and small-scale datasets (ImageNet, CIFAR-100, CIFAR-10), demonstrating improved uncertainty quantification capabilities compared to softmax MC sampling.
comment: 35 pages
♻ ☆ Conditional regression for the Nonlinear Single-Variable Model
Regressing a function $F$ on $\mathbb{R}^d$ without the statistical and computational curse of dimensionality requires special statistical models, for example that impose geometric assumptions on the distribution of the data (e.g., that its support is low-dimensional), or strong smoothness assumptions on $F$, or a special structure $F$. Among the latter, compositional models $F=f\circ g$ with $g$ mapping to $\mathbb{R}^r$ with $r\ll d$ include classical single- and multi-index models, as well as neural networks. While the case where $g$ is linear is well-understood, less is known when $g$ is nonlinear, and in particular for which $g$'s the curse of dimensionality in estimating $F$, or both $f$ and $g$, may be circumvented. Here we consider a model $F(X):=f(\Pi_\gamma X)$ where $\Pi_\gamma:\mathbb{R}^d\to[0,\textrm{len}_\gamma]$ is the closest-point projection onto the parameter of a regular curve $\gamma:[0, \textrm{len}_\gamma]\to\mathbb{R}^d$, and $f:[0,\textrm{len}_\gamma]\to \mathbb{R}^1$. The input data $X$ is not low-dimensional: it can be as far from $\gamma$ as the condition that $\Pi_\gamma(X)$ is well-defined allows. The distribution $X$, the curve $\gamma$ and the function $f$ are all unknown. This model is a natural nonlinear generalization of the single-index model, corresponding to $\gamma$ being a line. We propose a nonparametric estimator, based on conditional regression, that under suitable assumptions, the strongest of which being that $f$ is coarsely monotone, achieves, up to log factors, the $\textit{one-dimensional}$ optimal min-max rate for non-parametric regression, up to the level of noise in the observations, and be constructed in time $\mathcal{O}(d^2 n\log n)$. All the constants in the learning bounds, in the minimal number of samples required for our bounds to hold, and in the computational complexity are at most low-order polynomials in $d$.
comment: 57 pages, 10 figures
♻ ☆ SEREP: Semantic Facial Expression Representation for Robust In-the-Wild Capture and Retargeting
Monocular facial performance capture in-the-wild is challenging due to varied capture conditions, face shapes, and expressions. Most current methods rely on linear 3D Morphable Models, which represent facial expressions independently of identity at the vertex displacement level. We propose SEREP (Semantic Expression Representation), a model that disentangles expression from identity at the semantic level. We start by learning an expression representation from high-quality 3D data of unpaired facial expressions. Then, we train a model to predict expression from monocular images relying on a novel semi-supervised scheme using low quality synthetic data. In addition, we introduce MultiREX, a benchmark addressing the lack of evaluation resources for the expression capture task. Our experiments show that SEREP outperforms state-of-the-art methods, capturing challenging expressions and transferring them to new identities.
comment: For our project page, see https://ubisoft-laforge.github.io/character/serep/
♻ ☆ Learnable quantum spectral filters for hybrid graph neural networks
In this paper, we describe a parameterized quantum circuit that can be considered as convolutional and pooling layers for graph neural networks. The circuit incorporates the parameterized quantum Fourier circuit where the qubit connections for the controlled gates derived from the Laplacian operator. Specifically, we show that the eigenspace of the Laplacian operator of a graph can be approximated by using QFT based circuit whose connections are determined from the adjacency matrix. For an $N\times N$ Laplacian, this approach yields an approximate polynomial-depth circuit requiring only $n=log(N)$ qubits. These types of circuits can eliminate the expensive classical computations for approximating the learnable functions of the Laplacian through Chebyshev polynomial or Taylor expansions. Using this circuit as a convolutional layer provides an $n-$ dimensional probability vector that can be considered as the filtered and compressed graph signal. Therefore, the circuit along with the measurement can be considered a very efficient convolution plus pooling layer that transforms an $N$-dimensional signal input into $n-$dimensional signal with an exponential compression. We then apply a classical neural network prediction head to the output of the circuit to construct a complete graph neural network. Since the circuit incorporates geometric structure through its graph connection-based approach, we present graph classification results for the benchmark datasets listed in TUDataset library. Using only [1-100] learnable parameters for the quantum circuit and minimal classical layers (1000-5000 parameters) in a generic setting, the obtained results are comparable to and in some cases better than many of the baseline results, particularly for the cases when geometric structure plays a significant role.
comment: The simulation code and results used for this paper is publicly available at: https://github.com/adaskin/gnn-qsf
♻ ☆ Open Materials Generation with Stochastic Interpolants ICML
The discovery of new materials is essential for enabling technological advancements. Computational approaches for predicting novel materials must effectively learn the manifold of stable crystal structures within an infinite design space. We introduce Open Materials Generation (OMatG), a unifying framework for the generative design and discovery of inorganic crystalline materials. OMatG employs stochastic interpolants (SI) to bridge an arbitrary base distribution to the target distribution of inorganic crystals via a broad class of tunable stochastic processes, encompassing both diffusion models and flow matching as special cases. In this work, we adapt the SI framework by integrating an equivariant graph representation of crystal structures and extending it to account for periodic boundary conditions in unit cell representations. Additionally, we couple the SI flow over spatial coordinates and lattice vectors with discrete flow matching for atomic species. We benchmark OMatG's performance on two tasks: Crystal Structure Prediction (CSP) for specified compositions, and 'de novo' generation (DNG) aimed at discovering stable, novel, and unique structures. In our ground-up implementation of OMatG, we refine and extend both CSP and DNG metrics compared to previous works. OMatG establishes a new state of the art in generative modeling for materials discovery, outperforming purely flow-based and diffusion-based implementations. These results underscore the importance of designing flexible deep learning frameworks to accelerate progress in materials science. The OMatG code is available at https://github.com/FERMat-ML/OMatG.
comment: Accepted at Forty-second International Conference on Machine Learning (ICML): https://openreview.net/forum?id=gHGrzxFujU
♻ ☆ Feature Learning beyond the Lazy-Rich Dichotomy: Insights from Representational Geometry ICML 2025
Integrating task-relevant information into neural representations is a fundamental ability of both biological and artificial intelligence systems. Recent theories have categorized learning into two regimes: the rich regime, where neural networks actively learn task-relevant features, and the lazy regime, where networks behave like random feature models. Yet this simple lazy-rich dichotomy overlooks a diverse underlying taxonomy of feature learning, shaped by differences in learning algorithms, network architectures, and data properties. To address this gap, we introduce an analysis framework to study feature learning via the geometry of neural representations. Rather than inspecting individual learned features, we characterize how task-relevant representational manifolds evolve throughout the learning process. We show, in both theoretical and empirical settings, that as networks learn features, task-relevant manifolds untangle, with changes in manifold geometry revealing distinct learning stages and strategies beyond the lazy-rich dichotomy. This framework provides novel insights into feature learning across neuroscience and machine learning, shedding light on structural inductive biases in neural circuits and the mechanisms underlying out-of-distribution generalization.
comment: This work was published in ICML 2025 and was selected for a spotlight presentation
♻ ☆ Mind the Memory Gap: Unveiling GPU Bottlenecks in Large-Batch LLM Inference
Large language models have been widely adopted across different tasks, but their auto-regressive generation nature often leads to inefficient resource utilization during inference. While batching is commonly used to increase throughput, performance gains plateau beyond a certain batch size, especially with smaller models, a phenomenon that existing literature typically explains as a shift to the compute-bound regime. In this paper, through an in-depth GPU-level analysis, we reveal that large-batch inference remains memory-bound, with most GPU compute capabilities underutilized due to DRAM bandwidth saturation as the primary bottleneck. To address this, we propose a Batching Configuration Advisor (BCA) that optimizes memory allocation, reducing GPU memory requirements with minimal impact on throughput. The freed memory and underutilized GPU compute capabilities can then be leveraged by concurrent workloads. Specifically, we use model replication to improve serving throughput and GPU utilization. Our findings challenge conventional assumptions about LLM inference, offering new insights and practical strategies for improving resource utilization, particularly for smaller language models. The code is publicly available at https://github.com/FerranAgulloLopez/vLLMBatchingMemoryGap.
comment: Pol G. Recasens, Ferran Agullo: equal contribution. Paper accepted at IEEE CLOUD 2025
♻ ☆ A Malliavin calculus approach to score functions in diffusion generative models
Score-based diffusion generative models have recently emerged as a powerful tool for modelling complex data distributions. These models aim at learning the score function, which defines a map from a known probability distribution to the target data distribution via deterministic or stochastic differential equations (SDEs). The score function is typically estimated from data using a variety of approximation techniques, such as denoising or sliced score matching, Hyv\"arien's method, or Schr\"odinger bridges. In this paper, we derive an exact, closed form, expression for the score function for a broad class of nonlinear diffusion generative models. Our approach combines modern stochastic analysis tools such as Malliavin derivatives and their adjoint operators (Skorokhod integrals or Malliavin Divergence) with a new Bismut-type formula. The resulting expression for the score function can be written entirely in terms of the first and second variation processes, with all Malliavin derivatives systematically eliminated, thereby enhancing its practical applicability. The theoretical framework presented in this work offers a principled foundation for advancing score estimation methods in generative modelling, enabling the design of new sampling algorithms for complex probability distributions. Our results can be extended to broader classes of stochastic differential equations, opening new directions for the development of score-based diffusion generative models.
♻ ☆ Attribution assignment for deep-generative sequence models enables interpretability analysis using positive-only data
Generative machine learning models offer a powerful framework for therapeutic design by efficiently exploring large spaces of biological sequences enriched for desirable properties. Unlike supervised learning methods, which require both positive and negative labeled data, generative models such as LSTMs can be trained solely on positively labeled sequences, for example, high-affinity antibodies. This is particularly advantageous in biological settings where negative data are scarce, unreliable, or biologically ill-defined. However, the lack of attribution methods for generative models has hindered the ability to extract interpretable biological insights from such models. To address this gap, we developed Generative Attribution Metric Analysis (GAMA), an attribution method for autoregressive generative models based on Integrated Gradients. We assessed GAMA using synthetic datasets with known ground truths to characterize its statistical behavior and validate its ability to recover biologically relevant features. We further demonstrated the utility of GAMA by applying it to experimental antibody-antigen binding data. GAMA enables model interpretability and the validation of generative sequence design strategies without the need for negative training data.
♻ ☆ MedSegFactory: Text-Guided Generation of Medical Image-Mask Pairs
This paper presents MedSegFactory, a versatile medical synthesis framework that generates high-quality paired medical images and segmentation masks across modalities and tasks. It aims to serve as an unlimited data repository, supplying image-mask pairs to enhance existing segmentation tools. The core of MedSegFactory is a dual-stream diffusion model, where one stream synthesizes medical images and the other generates corresponding segmentation masks. To ensure precise alignment between image-mask pairs, we introduce Joint Cross-Attention (JCA), enabling a collaborative denoising paradigm by dynamic cross-conditioning between streams. This bidirectional interaction allows both representations to guide each other's generation, enhancing consistency between generated pairs. MedSegFactory unlocks on-demand generation of paired medical images and segmentation masks through user-defined prompts that specify the target labels, imaging modalities, anatomical regions, and pathological conditions, facilitating scalable and high-quality data generation. This new paradigm of medical image synthesis enables seamless integration into diverse medical imaging workflows, enhancing both efficiency and accuracy. Extensive experiments show that MedSegFactory generates data of superior quality and usability, achieving competitive or state-of-the-art performance in 2D and 3D segmentation tasks while addressing data scarcity and regulatory constraints.
comment: 12 pages, 8 figures, The project page can be accessed via https://jwmao1.github.io/MedSegFactory_web
♻ ☆ SAMO: A Lightweight Sharpness-Aware Approach for Multi-Task Optimization with Joint Global-Local Perturbation ICCV 2025
Multi-task learning (MTL) enables a joint model to capture commonalities across multiple tasks, reducing computation costs and improving data efficiency. However, a major challenge in MTL optimization is task conflicts, where the task gradients differ in direction or magnitude, limiting model performance compared to single-task counterparts. Sharpness-aware minimization (SAM) minimizes task loss while simultaneously reducing the sharpness of the loss landscape. Our empirical observations show that SAM effectively mitigates task conflicts in MTL. Motivated by these findings, we explore integrating SAM into MTL but face two key challenges. While both the average loss gradient and individual task gradients-referred to as global and local information-contribute to SAM, how to combine them remains unclear. Moreover, directly computing each task gradient introduces significant computational and memory overheads. To address these challenges, we propose SAMO, a lightweight \textbf{S}harpness-\textbf{A}ware \textbf{M}ulti-task \textbf{O}ptimization approach, that leverages a joint global-local perturbation. The local perturbations are approximated using only forward passes and are layerwise normalized to improve efficiency. Extensive experiments on a suite of multi-task benchmarks demonstrate both the effectiveness and efficiency of our method. Code is available at https://github.com/OptMN-Lab/SAMO.
comment: Accepted to ICCV 2025
♻ ☆ On the Gaussian process limit of Bayesian Additive Regression Trees
Bayesian Additive Regression Trees (BART) is a nonparametric Bayesian regression technique of rising fame. It is a sum-of-decision-trees model, and is in some sense the Bayesian version of boosting. In the limit of infinite trees, it becomes equivalent to Gaussian process (GP) regression. This limit is known but has not yet led to any useful analysis or application. For the first time, I derive and compute the exact BART prior covariance function. With it I implement the infinite trees limit of BART as GP regression. Through empirical tests, I show that this limit is worse than standard BART in a fixed configuration, but also that tuning its hyperparameters in the natural GP way makes it competitive with BART. The advantage of using a GP surrogate of BART is the analytical likelihood, which simplifies model building and sidesteps the complex BART MCMC algorithm. More generally, this study opens new ways to understand and develop BART and GP regression. The implementation of BART as GP is available in the Python package lsqfitgp.
comment: Check out the software at https://github.com/Gattocrucco/lsqfitgp
♻ ☆ What should a neuron aim for? Designing local objective functions based on information theory ICLR 2025
In modern deep neural networks, the learning dynamics of the individual neurons is often obscure, as the networks are trained via global optimization. Conversely, biological systems build on self-organized, local learning, achieving robustness and efficiency with limited global information. We here show how self-organization between individual artificial neurons can be achieved by designing abstract bio-inspired local learning goals. These goals are parameterized using a recent extension of information theory, Partial Information Decomposition (PID), which decomposes the information that a set of information sources holds about an outcome into unique, redundant and synergistic contributions. Our framework enables neurons to locally shape the integration of information from various input classes, i.e. feedforward, feedback, and lateral, by selecting which of the three inputs should contribute uniquely, redundantly or synergistically to the output. This selection is expressed as a weighted sum of PID terms, which, for a given problem, can be directly derived from intuitive reasoning or via numerical optimization, offering a window into understanding task-relevant local information processing. Achieving neuron-level interpretability while enabling strong performance using local learning, our work advances a principled information-theoretic foundation for local learning strategies.
comment: Presented as an oral at ICLR 2025. Conference version: https://openreview.net/forum?id=CLE09ESvul, 24 pages, 11 figures
♻ ☆ Data-driven system identification using quadratic embeddings of nonlinear dynamics
We propose a novel data-driven method called QENDy (Quadratic Embedding of Nonlinear Dynamics) that not only allows us to learn quadratic representations of highly nonlinear dynamical systems, but also to identify the governing equations. The approach is based on an embedding of the system into a higher-dimensional feature space in which the dynamics become quadratic. Just like SINDy (Sparse Identification of Nonlinear Dynamics), our method requires trajectory data, time derivatives for the training data points, which can also be estimated using finite difference approximations, and a set of preselected basis functions, called dictionary. We illustrate the efficacy and accuracy of QENDy with the aid of various benchmark problems and compare its performance with SINDy and a deep learning method for identifying quadratic embeddings. Furthermore, we analyze the convergence of QENDy and SINDy in the infinite data limit, highlight their similarities and main differences, and compare the quadratic embedding with linearization techniques based on the Koopman operator.
♻ ☆ GNN-ACLP: Graph Neural Networks Based Analog Circuit Link Prediction
Circuit link prediction identifying missing component connections from incomplete netlists is crucial in automating analog circuit design. However, existing methods face three main challenges: 1) Insufficient use of topological patterns in circuit graphs reduces prediction accuracy; 2) Data scarcity due to the complexity of annotations hinders model generalization; 3) Limited adaptability to various netlist formats. We propose GNN-ACLP, a Graph Neural Networks (GNNs) based framework featuring three innovations to tackle these challenges. First, we introduce the SEAL (Subgraphs, Embeddings, and Attributes for Link Prediction) framework and achieve port-level accuracy in circuit link prediction. Second, we propose Netlist Babel Fish, a netlist format conversion tool leveraging retrieval-augmented generation (RAG) with a large language model (LLM) to enhance the compatibility of netlist formats. Finally, we construct SpiceNetlist, a comprehensive dataset that contains 775 annotated circuits across 10 different component classes. Experiments demonstrate accuracy improvements of 16.08% on SpiceNetlist, 11.38% on Image2Net, and 16.01% on Masala-CHAI compared to the baseline in intra-dataset evaluation, while maintaining accuracy from 92.05% to 99.07% in cross-dataset evaluation, exhibiting robust feature transfer capabilities.
comment: Code and data will be made available on request. V3 Update: Add Ablation Study and Discussion; Improve Introduction; Optimize Figures; Add references
♻ ☆ Leveraging priors on distribution functions for multi-arm bandits
We introduce Dirichlet Process Posterior Sampling (DPPS), a Bayesian non-parametric algorithm for multi-arm bandits based on Dirichlet Process (DP) priors. Like Thompson-sampling, DPPS is a probability-matching algorithm, i.e., it plays an arm based on its posterior-probability of being optimal. Instead of assuming a parametric class for the reward generating distribution of each arm, and then putting a prior on the parameters, in DPPS the reward generating distribution is directly modeled using DP priors. DPPS provides a principled approach to incorporate prior belief about the bandit environment, and in the noninformative limit of the DP posteriors (i.e. Bayesian Bootstrap), we recover Non Parametric Thompson Sampling (NPTS), a popular non-parametric bandit algorithm, as a special case of DPPS. We employ stick-breaking representation of the DP priors, and show excellent empirical performance of DPPS in challenging synthetic and real world bandit environments. Finally, using an information-theoretic analysis, we show non-asymptotic optimality of DPPS in the Bayesian regret setup.
comment: Camera ready version -- Reinforcement Learning Journal, 2025
♻ ☆ Multiaccuracy and Multicalibration via Proxy Groups
As the use of predictive machine learning algorithms increases in high-stakes decision-making, it is imperative that these algorithms are fair across sensitive groups. However, measuring and enforcing fairness in real-world applications can be challenging due to the missing or incomplete sensitive group information. Proxy-sensitive attributes have been proposed as a practical and effective solution in these settings, but only for parity-based fairness notions. Knowing how to evaluate and control for fairness with missing sensitive group data for newer, different, and more flexible frameworks, such as multiaccuracy and multicalibration, remain unexplored. In this work, we address this gap by demonstrating that in the absence of sensitive group data, proxy-sensitive attributes can provably used to derive actionable upper bounds on the true multiaccuracy and multicalibration violations, providing insights into a predictive model's potential worst-case fairness violations. Additionally, we show that adjusting models to satisfy multiaccuracy and multicalibration across proxy-sensitive attributes can significantly mitigate these violations for the true, but unknown, sensitive groups. Through several experiments on real-world datasets, we illustrate that approximate multiaccuracy and multicalibration can be achieved even when sensitive group data is incomplete or unavailable.
♻ ☆ Binary and Ternary Quantization Can Enhance Feature Discrimination
Quantization is widely applied in machine learning to reduce computational and storage costs for both data and models. Considering that classification tasks are fundamental to the field, it is crucial to investigate how quantization impacts classification performance. Traditional research has focused on quantization errors, assuming that larger errors generally lead to lower classification accuracy. However, this assumption lacks a solid theoretical foundation and often contradicts empirical observations. For example, despite introducing significant errors, $\{0,1\}$-binary and $\{0, \pm1\}$-ternary quantized data have sometimes achieved classification accuracy comparable or even superior to full-precision data. To reasonably explain this phenomenon, a more accurate evaluation of classification performance is required. To achieve this, we propose a direct analysis of the feature discrimination of quantized data, instead of focusing on quantization errors. Our analysis reveals that both binary and ternary quantization can potentially enhance, rather than degrade, the feature discrimination of the original data. This finding is supported by classification experiments conducted on both synthetic and real data.
♻ ☆ Communities in the Kuramoto Model: Dynamics and Detection via Path Signatures
The behavior of multivariate dynamical processes is often governed by underlying structural connections that relate the components of the system. For example, brain activity, which is often measured via time series is determined by an underlying structural graph, where nodes represent neurons or brain regions and edges cortical connectivity. Existing methods for inferring structural connections from observed dynamics, such as correlation-based or spectral techniques, may fail to fully capture complex relationships in high-dimensional time series in an interpretable way. Here, we propose the use of path signatures, a mathematical framework that encodes geometric and temporal properties of continuous paths, to address this problem. Path signatures provide a reparametrization-invariant characterization of dynamical data and can be used to compute the lead matrix, which reveals lead-lag phenomena. We showcase our approach on time series from coupled oscillators in the Kuramoto model defined on a stochastic block model graph, termed the Kuramoto Stochastic Block Model (KSBM). Using mean-field theory and Gaussian approximations, we analytically derive reduced models of KSBM dynamics in different temporal regimes and theoretically characterize the lead matrix in these settings. Leveraging these insights, we propose a novel signature-based community detection algorithm, achieving exact recovery of structural communities from observed time series in multiple KSBM instances. We also explored the performance of our community detection on a stochastic variant of the KSBM as well as on real neuropixels of cortical recordings to demonstrate applicability on real-world data. Our results demonstrate that path signatures provide a novel perspective on analyzing complex neural data and other high-dimensional systems, explicitly exploiting temporal functional relationships to infer underlying structure.
comment: 56 pages, 21 figures
♻ ☆ REGEN: A Dataset and Benchmarks with Natural Language Critiques and Narratives
This paper introduces a novel dataset REGEN (Reviews Enhanced with GEnerative Narratives), designed to benchmark the conversational capabilities of recommender Large Language Models (LLMs), addressing the limitations of existing datasets that primarily focus on sequential item prediction. REGEN extends the Amazon Product Reviews dataset by inpainting two key natural language features: (1) user critiques, representing user "steering" queries that lead to the selection of a subsequent item, and (2) narratives, rich textual outputs associated with each recommended item taking into account prior context. The narratives include product endorsements, purchase explanations, and summaries of user preferences. Further, we establish an end-to-end modeling benchmark for the task of conversational recommendation, where models are trained to generate both recommendations and corresponding narratives conditioned on user history (items and critiques). For this joint task, we introduce a modeling framework LUMEN (LLM-based Unified Multi-task Model with Critiques, Recommendations, and Narratives) which uses an LLM as a backbone for critiquing, retrieval and generation. We also evaluate the dataset's quality using standard auto-rating techniques and benchmark it by training both traditional and LLM-based recommender models. Our results demonstrate that incorporating critiques enhances recommendation quality by enabling the recommender to learn language understanding and integrate it with recommendation signals. Furthermore, LLMs trained on our dataset effectively generate both recommendations and contextual narratives, achieving performance comparable to state-of-the-art recommenders and language models.
♻ ☆ One-Pass to Reason: Token Duplication and Block-Sparse Mask for Efficient Fine-Tuning on Multi-Turn Reasoning
Fine-tuning Large Language Models (LLMs) on multi-turn reasoning datasets requires N (number of turns) separate forward passes per conversation due to reasoning token visibility constraints, as reasoning tokens for a turn are discarded in subsequent turns. We propose duplicating response tokens along with a custom attention mask to enable single-pass processing of entire conversations. We prove our method produces identical losses to the N-pass approach while reducing time complexity from $O\bigl(N^{3}\bigl)$ to $O\bigl(N^{2}\bigl)$ and maintaining the same memory complexity for a transformer based model. Our approach achieves significant training speedup while preserving accuracy. Our implementation is available online (https://github.com/devrev/One-Pass-to-Reason).
comment: 9 pages, 3 figures
♻ ☆ Feasibility Study of CNNs and MLPs for Radiation Heat Transfer in 2-D Furnaces with Spectrally Participative Gases
Aiming to reduce the computational cost of numerical simulations, a convolutional neural network (CNN) and a multi-layer perceptron (MLP) are introduced to build a surrogate model to approximate radiative heat transfer solutions in a 2-D walled domain with participative gases. The originality of this work lays in the adaptation of the inputs of the problem (gas and wall properties) in order to fit with the CNN architecture, more commonly used for image processing. Two precision datasets have been created with the classical solver, ICARUS2D, that uses the discrete transfer radiation method with the statistical narrow bands model. The performance of the CNN architecture is compared to a more classical MLP architecture in terms of speed and accuracy. Thanks to Optuna, all results are obtained using the optimized hyper parameters networks. The results show a significant speedup with industrially acceptable relative errors compared to the classical solver for both architectures. Additionally, the CNN outperforms the MLP in terms of precision and is more robust and stable to changes in hyper-parameters. A performance analysis on the dataset size of the samples have also been carried out to gain a deeper understanding of the model behavior.
♻ ☆ Squeeze the Soaked Sponge: Efficient Off-policy Reinforcement Finetuning for Large Language Model
Reinforcement Learning (RL) has demonstrated its potential to improve the reasoning ability of Large Language Models (LLMs). One major limitation of most existing Reinforcement Finetuning (RFT) methods is that they are on-policy RL in nature, i.e., data generated during the past learning process is not fully utilized. This inevitably comes at a significant cost of compute and time, posing a stringent bottleneck on continuing economic and efficient scaling. To this end, we launch the renaissance of off-policy RL and propose Reincarnating Mix-policy Proximal Policy Gradient (ReMix), a general approach to enable on-policy RFT methods like PPO and GRPO to leverage off-policy data. ReMix consists of three major components: (1) Mix-policy proximal policy gradient with an increased Update-To-Data (UTD) ratio for efficient training; (2) KL-Convex policy constraint to balance the trade-off between stability and flexibility; (3) Policy reincarnation to achieve a seamless transition from efficient early-stage learning to steady asymptotic improvement. In our experiments, we train a series of ReMix models upon PPO, GRPO and 1.5B, 7B base models. ReMix shows an average Pass@1 accuracy of 52.10% (for 1.5B model) with 0.079M response rollouts, 350 training steps and achieves 63.27%/64.39% (for 7B model) with 0.007M/0.011M response rollouts, 50/75 training steps, on five math reasoning benchmarks (i.e., AIME'24, AMC'23, Minerva, OlympiadBench, and MATH500). Compared with 15 recent advanced models, ReMix shows SOTA-level performance with an over 30x to 450x reduction in training cost in terms of rollout data volume. In addition, we reveal insightful findings via multifaceted analysis, including the implicit preference for shorter responses due to the Whipping Effect of off-policy discrepancy, the collapse mode of self-reflection behavior under the presence of severe off-policyness, etc.
comment: Preliminary version, v3, added the missing name of x-axis in the left part of Fig.1 and corrected a wrong number in Fig.3. Project page: https://anitaleungxx.github.io/ReMix
♻ ☆ Predicting Air Pollution in Cork, Ireland Using Machine Learning
Air pollution poses a critical health threat in cities worldwide, with nitrogen dioxide levels in Cork, Ireland exceeding World Health Organization safety standards by up to $278\%$. This study leverages artificial intelligence to predict air pollution with unprecedented accuracy, analyzing nearly ten years of data from five monitoring stations combined with 30 years of weather records. We evaluated 17 machine learning algorithms, with Extra Trees emerging as the optimal solution, achieving $77\%$ prediction accuracy and significantly outperforming traditional forecasting methods. Our analysis reveals that meteorological conditions particularly temperature, wind speed, and humidity are the primary drivers of pollution levels, while traffic patterns and seasonal changes create predictable pollution cycles. Pollution exhibits dramatic seasonal variations, with winter levels nearly double those of summer, and daily rush-hour peaks reaching $120\%$ above normal levels. While Cork's air quality shows concerning violations of global health standards, our models detected an encouraging $31\%$ improvement from 2014 to 2022. This research demonstrates that intelligent forecasting systems can provide city planners and environmental officials with powerful prediction tools, enabling life-saving early warning systems and informed urban planning decisions. The technology exists today to transform urban air quality management. All research materials and code are freely available at: https://github.com/MdRashidunnabi/Air-Pollution-Analysis.git
comment: The draft was submitted prematurely and requires further analysis, added research findings, and corrected references. Some co-authors have not yet approved this version. I will ensure all necessary revisions and approvals before resubmitting
♻ ☆ Neural Concept Verifier: Scaling Prover-Verifier Games via Concept Encodings
While Prover-Verifier Games (PVGs) offer a promising path toward verifiability in nonlinear classification models, they have not yet been applied to complex inputs such as high-dimensional images. Conversely, Concept Bottleneck Models (CBMs) effectively translate such data into interpretable concepts but are limited by their reliance on low-capacity linear predictors. In this work, we introduce the Neural Concept Verifier (NCV), a unified framework combining PVGs with concept encodings for interpretable, nonlinear classification in high-dimensional settings. NCV achieves this by utilizing recent minimally supervised concept discovery models to extract structured concept encodings from raw inputs. A prover then selects a subset of these encodings, which a verifier -- implemented as a nonlinear predictor -- uses exclusively for decision-making. Our evaluations show that NCV outperforms CBM and pixel-based PVG classifier baselines on high-dimensional, logically complex datasets and also helps mitigate shortcut behavior. Overall, we demonstrate NCV as a promising step toward performative, verifiable AI.
comment: 16 pages, 4 figures, 8 tables, revised references
♻ ☆ Sculpting Quantum Landscapes: Fubini-Study Metric Conditioning for Geometry Aware Learning in Parameterized Quantum Circuits
We present a novel meta learning framework called Sculpture that explicitly conditions the Fubini Study metric tensor of parameterized quantum circuits to mitigate barren plateaus in variational quantum algorithms. Our theoretical analysis identifies the logarithmic condition number of the Fubini Study metric as a critical geometric quantity governing trainability, optimization dynamics, and generalization. Sculpture uses a classical meta model trained to generate data dependent quantum circuit initializations that minimize the logarithmic condition number, thereby promoting an isotropic and well conditioned parameter space. Empirical results show that meta training reduces the logarithmic condition number from approximately 1.47 to 0.64 by significantly increasing the minimum eigenvalue and slightly decreasing the maximum eigenvalue of the metric, effectively alleviating barren plateaus. This improved conditioning generalizes well to unseen data, consistently producing well conditioned quantum circuit initializations. In a downstream hybrid quantum classical classification task on the Kaggle diabetes dataset, increasing the meta scaling coefficient accelerates convergence, reduces training loss and gradient norms, and crucially improves generalization, with test accuracy increasing from about 0.68 to over 0.78. These findings demonstrate that sculpting the quantum landscape via meta learning serves as a principled geometric regularizer, substantially enhancing trainability, optimization, and generalization of parameterized quantum circuits and enabling more robust and efficient variational quantum algorithms.
comment: Need more analysis
♻ ☆ Collaborative filtering based on nonnegative/binary matrix factorization
Collaborative filtering generates recommendations by exploiting user-item similarities based on rating data, which often contains numerous unrated items. This paper proposes a nonnegative/binary matrix factorization (NBMF) algorithm modified for collaborative filtering and demonstrates that utilizing a low-latency Ising machine in NBMF is advantageous in terms of computation time. While previous studies have primarily applied NBMF to dense data, such as images, this study applies a modified NBMF to sparse data. Results show the benefits of using a low-latency Ising machine to implement the proposed method.
comment: 12 pages, 8 figures
♻ ☆ Field Matching: an Electrostatic Paradigm to Generate and Transfer Data
We propose Electrostatic Field Matching (EFM), a novel method that is suitable for both generative modeling and distribution transfer tasks. Our approach is inspired by the physics of an electrical capacitor. We place source and target distributions on the capacitor plates and assign them positive and negative charges, respectively. We then learn the electrostatic field of the capacitor using a neural network approximator. To map the distributions to each other, we start at one plate of the capacitor and move the samples along the learned electrostatic field lines until they reach the other plate. We theoretically justify that this approach provably yields the distribution transfer. In practice, we demonstrate the performance of our EFM in toy and image data experiments.
comment: Proceedings of the 42nd International Conference on Machine. Learning, Vancouver, Canada. PMLR 267, 2025
♻ ☆ Average Calibration Error: A Differentiable Loss for Improved Reliability in Image Segmentation
Deep neural networks for medical image segmentation often produce overconfident results misaligned with empirical observations. Such miscalibration, challenges their clinical translation. We propose to use marginal L1 average calibration error (mL1-ACE) as a novel auxiliary loss function to improve pixel-wise calibration without compromising segmentation quality. We show that this loss, despite using hard binning, is directly differentiable, bypassing the need for approximate but differentiable surrogate or soft binning approaches. Our work also introduces the concept of dataset reliability histograms which generalises standard reliability diagrams for refined visual assessment of calibration in semantic segmentation aggregated at the dataset level. Using mL1-ACE, we reduce average and maximum calibration error by 45% and 55% respectively, maintaining a Dice score of 87% on the BraTS 2021 dataset. We share our code here: https://github.com/cai4cai/ACE-DLIRIS
comment: accidental replacement intended for arXiv:2506.03942
♻ ☆ FonTS: Text Rendering with Typography and Style Controls ICCV 2025
Visual text rendering are widespread in various real-world applications, requiring careful font selection and typographic choices. Recent progress in diffusion transformer (DiT)-based text-to-image (T2I) models show promise in automating these processes. However, these methods still encounter challenges like inconsistent fonts, style variation, and limited fine-grained control, particularly at the word-level. This paper proposes a two-stage DiT-based pipeline to address these problems by enhancing controllability over typography and style in text rendering. We introduce typography control fine-tuning (TC-FT), an parameter-efficient fine-tuning method (on $5\%$ key parameters) with enclosing typography control tokens (ETC-tokens), which enables precise word-level application of typographic features. To further address style inconsistency in text rendering, we propose a text-agnostic style control adapter (SCA) that prevents content leakage while enhancing style consistency. To implement TC-FT and SCA effectively, we incorporated HTML-render into the data synthesis pipeline and proposed the first word-level controllable dataset. Through comprehensive experiments, we demonstrate the effectiveness of our approach in achieving superior word-level typographic control, font consistency, and style consistency in text rendering tasks. The datasets and models will be available for academic use.
comment: Accepted to ICCV 2025
♻ ☆ Answer Generation for Questions With Multiple Information Sources in E-Commerce
Automatic question answering is an important yet challenging task in E-commerce given the millions of questions posted by users about the product that they are interested in purchasing. Hence, there is a great demand for automatic answer generation systems that provide quick responses using related information about the product. There are three sources of knowledge available for answering a user posted query, they are reviews, duplicate or similar questions, and specifications. Effectively utilizing these information sources will greatly aid us in answering complex questions. However, there are two main challenges present in exploiting these sources: (i) The presence of irrelevant information and (ii) the presence of ambiguity of sentiment present in reviews and similar questions. Through this work we propose a novel pipeline (MSQAP) that utilizes the rich information present in the aforementioned sources by separately performing relevancy and ambiguity prediction before generating a response. Experimental results show that our relevancy prediction model (BERT-QA) outperforms all other variants and has an improvement of 12.36% in F1 score compared to the BERT-base baseline. Our generation model (T5-QA) outperforms the baselines in all content preservation metrics such as BLEU, ROUGE and has an average improvement of 35.02% in ROUGE and 198.75% in BLEU compared to the highest performing baseline (HSSC-q). Human evaluation of our pipeline shows us that our method has an overall improvement in accuracy of 30.7% over the generation model (T5-QA), resulting in our full pipeline-based approach (MSQAP) providing more accurate answers. To the best of our knowledge, this is the first work in the e-commerce domain that automatically generates natural language answers combining the information present in diverse sources such as specifications, similar questions, and reviews data.
comment: 7 pages, 10 tables, 1 figure
♻ ☆ Minerva: A File-Based Ransomware Detector CCS 2025
Ransomware attacks have caused billions of dollars in damages in recent years, and are expected to cause billions more in the future. Consequently, significant effort has been devoted to ransomware detection and mitigation. Behavioral-based ransomware detection approaches have garnered considerable attention recently. These behavioral detectors typically rely on process-based behavioral profiles to identify malicious behaviors. However, with an increasing body of literature highlighting the vulnerability of such approaches to evasion attacks, a comprehensive solution to the ransomware problem remains elusive. This paper presents Minerva, a novel, robust approach to ransomware detection. Minerva is engineered to be robust by design against evasion attacks, with architectural and feature selection choices informed by their resilience to adversarial manipulation. We conduct a comprehensive analysis of Minerva across a diverse spectrum of ransomware types, encompassing unseen ransomware as well as variants designed specifically to evade Minerva. Our evaluation showcases the ability of Minerva to accurately identify ransomware, generalize to unseen threats, and withstand evasion attacks. Furthermore, over 99% of detected ransomware are identified within 0.52sec of activity, enabling the adoption of data loss prevention techniques with near-zero overhead.
comment: Accepted for publication at The 20th ACM ASIA Conference on Computer and Communications Security (ACM ASIACCS 2025), Meli\'a Hanoi
♻ ☆ DriveTransformer: Unified Transformer for Scalable End-to-End Autonomous Driving ICLR2025
End-to-end autonomous driving (E2E-AD) has emerged as a trend in the field of autonomous driving, promising a data-driven, scalable approach to system design. However, existing E2E-AD methods usually adopt the sequential paradigm of perception-prediction-planning, which leads to cumulative errors and training instability. The manual ordering of tasks also limits the system`s ability to leverage synergies between tasks (for example, planning-aware perception and game-theoretic interactive prediction and planning). Moreover, the dense BEV representation adopted by existing methods brings computational challenges for long-range perception and long-term temporal fusion. To address these challenges, we present DriveTransformer, a simplified E2E-AD framework for the ease of scaling up, characterized by three key features: Task Parallelism (All agent, map, and planning queries direct interact with each other at each block), Sparse Representation (Task queries direct interact with raw sensor features), and Streaming Processing (Task queries are stored and passed as history information). As a result, the new framework is composed of three unified operations: task self-attention, sensor cross-attention, temporal cross-attention, which significantly reduces the complexity of system and leads to better training stability. DriveTransformer achieves state-of-the-art performance in both simulated closed-loop benchmark Bench2Drive and real world open-loop benchmark nuScenes with high FPS.
comment: Accepted by ICLR2025; Fix Typo
♻ ☆ Estimation of conditional average treatment effects on distributed confidential data
The estimation of conditional average treatment effects (CATEs) is an important topic in many scientific fields. CATEs can be estimated with high accuracy if data distributed across multiple parties are centralized. However, it is difficult to aggregate such data owing to confidentiality or privacy concerns. To address this issue, we propose data collaboration double machine learning, a method for estimating CATE models using privacy-preserving fusion data constructed from distributed sources, and evaluate its performance through simulations. We make three main contributions. First, our method enables estimation and testing of semi-parametric CATE models without iterative communication on distributed data, providing robustness to model mis-specification compared to parametric approaches. Second, it enables collaborative estimation across different time points and parties by accumulating a knowledge base. Third, our method performs as well as or better than existing methods in simulations using synthetic, semi-synthetic, and real-world datasets.
comment: 45 pages, 12 figures
♻ ☆ Sampling from Your Language Model One Byte at a Time
Tokenization is used almost universally by modern language models, enabling efficient text representation using multi-byte or multi-character tokens. However, prior work has shown that tokenization can introduce distortion into the model's generations, an issue known as the Prompt Boundary Problem (PBP). For example, users are often advised not to end their prompts with a space because it prevents the model from including the space as part of the next token. While this heuristic is effective in English, the underlying PBP continues to affect languages such as Chinese as well as code generation, where tokens often do not line up with word and syntactic boundaries. In this work, we present an inference-time method to convert any autoregressive LM with a BPE tokenizer into a character-level or byte-level LM. Our method efficiently solves the PBP and is also able to unify the vocabularies of language models with different tokenizers, allowing one to ensemble LMs with different tokenizers at inference time or transfer the post-training from one model to another using proxy-tuning. We demonstrate in experiments that the ensemble and proxy-tuned models outperform their constituents on downstream evals. Code is available at https://github.com/SewoongLab/byte-sampler .
comment: 23 pages, 8 figures
♻ ☆ Learning Pole Structures of Hadronic States using Predictive Uncertainty Estimation
Matching theoretical predictions to experimental data remains a central challenge in hadron spectroscopy. In particular, the identification of new hadronic states is difficult, as exotic signals near threshold can arise from a variety of physical mechanisms. A key diagnostic in this context is the pole structure of the scattering amplitude, but different configurations can produce similar signatures. The mapping between pole configurations and line shapes is especially ambiguous near the mass threshold, where analytic control is limited. In this work, we introduce an uncertainty-aware machine learning approach for classifying pole structures in $S$-matrix elements. Our method is based on an ensemble of classifier chains that provide both epistemic and aleatoric uncertainty estimates. We apply a rejection criterion based on predictive uncertainty, achieving a validation accuracy of nearly $95\%$ while discarding only a small fraction of high-uncertainty predictions. Trained on synthetic data with known pole structures, the model generalizes to previously unseen experimental data, including enhancements associated with the $P_{c\bar{c}}(4312)^+$ state observed by LHCb. In this, we infer a four-pole structure, representing the presence of a genuine compact pentaquark in the presence of a higher channel virtual state pole with non-vanishing width. While evaluated on this particular state, our framework is broadly applicable to other candidate hadronic states and offers a scalable tool for pole structure inference in scattering amplitudes.
♻ ☆ A Plea for History and Philosophy of Statistics and Machine Learning
The integration of the history and philosophy of statistics was initiated at least by Hacking (1965) and advanced by Mayo (1996), but it has not received sustained follow-up. Yet such integration is more urgent than ever, as the recent success of artificial intelligence has been driven largely by machine learning -- a field historically developed alongside statistics. Today, the boundary between statistics and machine learning is increasingly blurred. What we now need is integration, twice over: of history and philosophy, and of two fields they engage -- statistics and machine learning. I present a case study of a philosophical idea in machine learning (and in formal epistemology) whose root can be traced back to an often under-appreciated insight in Neyman and Pearson's 1936 work (a follow-up to their 1933 classic). This leads to the articulation of an epistemological principle -- largely implicit in, but shared by, the practices of frequentist statistics and machine learning -- which I call achievabilism: the thesis that the correct standard for assessing non-deductive inference methods should not be fixed, but should instead be sensitive to what is achievable in specific problem contexts. Another integration also emerges at the level of methodology, combining two ends of the philosophy of science spectrum: history and philosophy of science on the one hand, and formal epistemology on the other hand.
♻ ☆ DRAN: A Distribution and Relation Adaptive Network for Spatio-temporal Forecasting
Accurate predictions of spatio-temporal systems are crucial for tasks such as system management, control, and crisis prevention. However, the inherent time variance of many spatio-temporal systems poses challenges to achieving accurate predictions whenever stationarity is not granted. In order to address non-stationarity, we propose a Distribution and Relation Adaptive Network (DRAN) capable of dynamically adapting to relation and distribution changes over time. While temporal normalization and de-normalization are frequently used techniques to adapt to distribution shifts, this operation is not suitable for the spatio-temporal context as temporal normalization scales the time series of nodes and possibly disrupts the spatial relations among nodes. In order to address this problem, a Spatial Factor Learner (SFL) module is developed that enables the normalization and de-normalization process. To adapt to dynamic changes in spatial relationships among sensors, we propose a Dynamic-Static Fusion Learner (DSFL) module that effectively integrates features learned from both dynamic and static relations through an adaptive fusion ratio mechanism. Furthermore, we introduce a Stochastic Learner to capture the noisy components of spatio-temporal representations. Our approach outperforms state-of-the-art methods on weather prediction and traffic flow forecasting tasks.Experimental results show that our SFL efficiently preserves spatial relationships across various temporal normalization operations. Visualizations of the learned dynamic and static relations demonstrate that DSFL can capture both local and distant relationships between nodes.
comment: 15 pages, 10 figures
♻ ☆ Galerkin-ARIMA: A Two-Stage Polynomial Regression Framework for Fast Rolling One-Step-Ahead Forecasting
We introduce Galerkin-ARIMA, a novel time-series forecasting framework that integrates Galerkin projection techniques with the classical ARIMA model to capture potentially nonlinear dependencies in lagged observations. By replacing the fixed linear autoregressive component with a spline-based basis expansion, Galerkin-ARIMA flexibly approximates the underlying relationship among past values via ordinary least squares, while retaining the moving-average structure and Gaussian innovation assumptions of ARIMA. We derive closed-form solutions for both the AR and MA components using two-stage Galerkin projections, establish conditions for asymptotic unbiasedness and consistency, and analyze the bias-variance trade-off under basis-size growth. Complexity analysis reveals that, for moderate basis dimensions, our approach can substantially reduce computational cost compared to maximum-likelihood ARIMA estimation. Through extensive simulations on four synthetic processes-including noisy ARMA, seasonal, trend-AR, and nonlinear recursion series-we demonstrate that Galerkin-ARIMA matches or closely approximates ARIMA's forecasting accuracy while achieving orders-of-magnitude speedups in rolling forecasting tasks. These results suggest that Galerkin-ARIMA offers a powerful, efficient alternative for modeling complex time series dynamics in high-volume or real-time applications.
♻ ☆ Enhancing Distributional Robustness in Principal Component Analysis by Wasserstein Distances
We consider the distributionally robust optimization (DRO) model of principal component analysis (PCA) to account for uncertainty in the underlying probability distribution. The resulting formulation leads to a nonsmooth constrained min-max optimization problem, where the ambiguity set captures the distributional uncertainty by the type-$2$ Wasserstein distance. We prove that the inner maximization problem admits a closed-form optimal value. This explicit characterization equivalently reformulates the original DRO model into a minimization problem on the Stiefel manifold with intricate nonsmooth terms, a challenging formulation beyond the reach of existing algorithms. To address this issue, we devise an efficient smoothing manifold proximal gradient algorithm. Our analysis establishes Riemannian gradient consistency and global convergence of our algorithm to a stationary point of the nonsmooth minimization problem. We also provide the iteration complexity $O(\epsilon^{-3})$ of our algorithm to achieve an $\epsilon$-approximate stationary point. Finally, numerical experiments are conducted to validate the effectiveness and scalability of our algorithm, as well as to highlight the necessity and rationality of adopting the DRO model for PCA.
♻ ☆ Emoji Attack: Enhancing Jailbreak Attacks Against Judge LLM Detection
Jailbreaking techniques trick Large Language Models (LLMs) into producing restricted output, posing a potential threat. One line of defense is to use another LLM as a Judge to evaluate the harmfulness of generated text. However, we reveal that these Judge LLMs are vulnerable to token segmentation bias, an issue that arises when delimiters alter the tokenization process, splitting words into smaller sub-tokens. This alters the embeddings of the entire sequence, reducing detection accuracy and allowing harmful content to be misclassified as safe. In this paper, we introduce Emoji Attack, a novel strategy that amplifies existing jailbreak prompts by exploiting token segmentation bias. Our method leverages in-context learning to systematically insert emojis into text before it is evaluated by a Judge LLM, inducing embedding distortions that significantly lower the likelihood of detecting unsafe content. Unlike traditional delimiters, emojis also introduce semantic ambiguity, making them particularly effective in this attack. Through experiments on state-of-the-art Judge LLMs, we demonstrate that Emoji Attack substantially reduces the unsafe prediction rate, bypassing existing safeguards.
♻ ☆ EvalTree: Profiling Language Model Weaknesses via Hierarchical Capability Trees
An ideal model evaluation should achieve two goals: identifying where the model fails and providing actionable improvement guidance. Toward these goals for language model (LM) evaluations, we formulate the problem of generating a weakness profile, a set of weaknesses expressed in natural language, given an LM's performance on every individual instance in a benchmark. We introduce a suite of quantitative assessments to compare different weakness profiling methods. We also introduce a weakness profiling method EvalTree. EvalTree constructs a capability tree where each node represents a capability described in natural language and is linked to a subset of benchmark instances that specifically evaluate this capability; it then extracts nodes where the LM performs poorly to generate a weakness profile. On the MATH and WildChat benchmarks, we show that EvalTree outperforms baseline weakness profiling methods by identifying weaknesses more precisely and comprehensively. Weakness profiling further enables weakness-guided data collection, and training data collection guided by EvalTree-identified weaknesses improves LM performance more than other data collection strategies. We also show how EvalTree exposes flaws in Chatbot Arena's human-voter-based evaluation practice. To facilitate future work, we provide an interface that allows practitioners to interactively explore the capability trees built by EvalTree.
comment: COLM 2025
♻ ☆ Unraveling the Interplay between Carryover Effects and Reward Autocorrelations in Switchback Experiments
A/B testing has become the gold standard for policy evaluation in modern technological industries. Motivated by the widespread use of switchback experiments in A/B testing, this paper conducts a comprehensive comparative analysis of various switchback designs in Markovian environments. Unlike many existing works which derive the optimal design based on specific and relatively simple estimators, our analysis covers a range of state-of-the-art estimators developed in the reinforcement learning (RL) literature. It reveals that the effectiveness of different switchback designs depends crucially on (i) the size of the carryover effect and (ii) the auto-correlations among reward errors over time. Meanwhile, these findings are estimator-agnostic, i.e., they apply to most RL estimators. Based on these insights, we provide a workflow to offer guidelines for practitioners on designing switchback experiments in A/B testing.
♻ ☆ Amortized Posterior Sampling with Diffusion Prior Distillation
We propose Amortized Posterior Sampling (APS), a novel variational inference approach for efficient posterior sampling in inverse problems. Our method trains a conditional flow model to minimize the divergence between the variational distribution and the posterior distribution implicitly defined by the diffusion model. This results in a powerful, amortized sampler capable of generating diverse posterior samples with a single neural function evaluation, generalizing across various measurements. Unlike existing methods, our approach is unsupervised, requires no paired training data, and is applicable to both Euclidean and non-Euclidean domains. We demonstrate its effectiveness on a range of tasks, including image restoration, manifold signal reconstruction, and climate data imputation. APS significantly outperforms existing approaches in computational efficiency while maintaining competitive reconstruction quality, enabling real-time, high-quality solutions to inverse problems across diverse domains.
♻ ☆ Bandit-Based Prompt Design Strategy Selection Improves Prompt Optimizers ACL 2025
Prompt optimization aims to search for effective prompts that enhance the performance of large language models (LLMs). Although existing prompt optimization methods have discovered effective prompts, they often differ from sophisticated prompts carefully designed by human experts. Prompt design strategies, representing best practices for improving prompt performance, can be key to improving prompt optimization. Recently, a method termed the Autonomous Prompt Engineering Toolbox (APET) has incorporated various prompt design strategies into the prompt optimization process. In APET, the LLM is needed to implicitly select and apply the appropriate strategies because prompt design strategies can have negative effects. This implicit selection may be suboptimal due to the limited optimization capabilities of LLMs. This paper introduces Optimizing Prompts with sTrategy Selection (OPTS), which implements explicit selection mechanisms for prompt design. We propose three mechanisms, including a Thompson sampling-based approach, and integrate them into EvoPrompt, a well-known prompt optimizer. Experiments optimizing prompts for two LLMs, Llama-3-8B-Instruct and GPT-4o mini, were conducted using BIG-Bench Hard. Our results show that the selection of prompt design strategies improves the performance of EvoPrompt, and the Thompson sampling-based mechanism achieves the best overall results. Our experimental code is provided at https://github.com/shiralab/OPTS .
comment: Accepted to ACL 2025 Findings
♻ ☆ Dualformer: Controllable Fast and Slow Thinking by Learning with Randomized Reasoning Traces
In cognition theory, human thinking is governed by two systems: the fast and intuitive System 1 and the slower but more deliberative System 2. Analogously, Large Language Models (LLMs) can operate in two reasoning modes: outputting only the solutions (\emph{fast mode}) or both the reasoning chain and the final solution (\emph{slow mode}). We present \dualformer, a single Transformer model that seamlessly integrates both the fast and slow reasoning modes by training on randomized reasoning traces, where different parts of the traces are strategically dropped during training. At inference time, \dualformer can be easily configured to execute in either fast or slow mode, or automatically decide which mode to engage (\emph{auto mode}). It outperforms baselines in both performance and computational efficiency across all three modes: (1) in slow mode, \dualformer achieves $97.6\%$ optimal rate on unseen $30 \times 30$ maze tasks, surpassing the \searchformer baseline ($93.3\%$) trained on data with complete reasoning traces, with $45.5\%$ fewer reasoning steps; (2) in fast mode, \dualformer achieves $80\%$ optimal rate, significantly outperforming the Solution-Only model trained on solution-only data, which has an optimal rate of only $30\%$; (3) in auto mode, \dualformer achieves $96.6\%$ optimal rate with $59.9\%$ fewer steps than \searchformer. Moreover, \dualformer produces more diverse reasoning traces than \searchformer{}. For math reasoning problems, our techniques have also achieved improved performance with LLM fine-tuning, demonstrating its generalization beyond task-specific models. We open source our code at https://github.com/facebookresearch/dualformer.
♻ ☆ Granular Ball Twin Support Vector Machine
On Efficient and Scalable Computation of the Nonparametric Maximum Likelihood Estimator in Mixture ModelsTwin support vector machine (TSVM) is an emerging machine learning model with versatile applicability in classification and regression endeavors. Nevertheless, TSVM confronts noteworthy challenges: $(i)$ the imperative demand for matrix inversions presents formidable obstacles to its efficiency and applicability on large-scale datasets; $(ii)$ the omission of the structural risk minimization (SRM) principle in its primal formulation heightens the vulnerability to overfitting risks; and $(iii)$ the TSVM exhibits a high susceptibility to noise and outliers, and also demonstrates instability when subjected to resampling. In view of the aforementioned challenges, we propose the granular ball twin support vector machine (GBTSVM). GBTSVM takes granular balls, rather than individual data points, as inputs to construct a classifier. These granular balls, characterized by their coarser granularity, exhibit robustness to resampling and reduced susceptibility to the impact of noise and outliers. We further propose a novel large-scale granular ball twin support vector machine (LS-GBTSVM). LS-GBTSVM's optimization formulation ensures two critical facets: $(i)$ it eliminates the need for matrix inversions, streamlining the LS-GBTSVM's computational efficiency, and $(ii)$ it incorporates the SRM principle through the incorporation of regularization terms, effectively addressing the issue of overfitting. The proposed LS-GBTSVM exemplifies efficiency, scalability for large datasets, and robustness against noise and outliers. We conduct a comprehensive evaluation of the GBTSVM and LS-GBTSVM models on benchmark datasets from UCI, KEEL, and NDC datasets. Our experimental findings and statistical analyses affirm the superior generalization prowess of the proposed GBTSVM and LS-GBTSVM models.
comment: Manuscript submitted to IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS: 19 September 2023; revised 13 February 2024 and 14 July 2024; accepted 05 October 2024
♻ ☆ Distributional Soft Actor-Critic with Diffusion Policy SC 2025
Reinforcement learning has been proven to be highly effective in handling complex control tasks. Traditional methods typically use unimodal distributions, such as Gaussian distributions, to model the output of value distributions. However, unimodal distribution often and easily causes bias in value function estimation, leading to poor algorithm performance. This paper proposes a distributional reinforcement learning algorithm called DSAC-D (Distributed Soft Actor Critic with Diffusion Policy) to address the challenges of estimating bias in value functions and obtaining multimodal policy representations. A multimodal distributional policy iteration framework that can converge to the optimal policy was established by introducing policy entropy and value distribution function. A diffusion value network that can accurately characterize the distribution of multi peaks was constructed by generating a set of reward samples through reverse sampling using a diffusion model. Based on this, a distributional reinforcement learning algorithm with dual diffusion of the value network and the policy network was derived. MuJoCo testing tasks demonstrate that the proposed algorithm not only learns multimodal policy, but also achieves state-of-the-art (SOTA) performance in all 9 control tasks, with significant suppression of estimation bias and total average return improvement of over 10% compared to existing mainstream algorithms. The results of real vehicle testing show that DSAC-D can accurately characterize the multimodal distribution of different driving styles, and the diffusion policy network can characterize multimodal trajectories.
comment: Accepted IEEE ITSC 2025
♻ ☆ Pocket2Mol: Efficient Molecular Sampling Based on 3D Protein Pockets ICML 2022
Deep generative models have achieved tremendous success in designing novel drug molecules in recent years. A new thread of works have shown the great potential in advancing the specificity and success rate of in silico drug design by considering the structure of protein pockets. This setting posts fundamental computational challenges in sampling new chemical compounds that could satisfy multiple geometrical constraints imposed by pockets. Previous sampling algorithms either sample in the graph space or only consider the 3D coordinates of atoms while ignoring other detailed chemical structures such as bond types and functional groups. To address the challenge, we develop Pocket2Mol, an E(3)-equivariant generative network composed of two modules: 1) a new graph neural network capturing both spatial and bonding relationships between atoms of the binding pockets and 2) a new efficient algorithm which samples new drug candidates conditioned on the pocket representations from a tractable distribution without relying on MCMC. Experimental results demonstrate that molecules sampled from Pocket2Mol achieve significantly better binding affinity and other drug properties such as druglikeness and synthetic accessibility.
comment: ICML 2022 accepted
♻ ☆ Local transfer learning Gaussian process modeling, with applications to surrogate modeling of expensive computer simulators
A critical bottleneck for scientific progress is the costly nature of computer simulations for complex systems. Surrogate models provide an appealing solution: such models are trained on simulator evaluations, then used to emulate and quantify uncertainty on the expensive simulator at unexplored inputs. In many applications, one often has available data on related systems. For example, in designing a new jet turbine, there may be existing studies on turbines with similar configurations. A key question is how information from such ``source'' systems can be transferred for effective surrogate training on the ``target'' system of interest. We thus propose a new LOcal transfer Learning Gaussian Process (LOL-GP) model, which leverages a carefully-designed Gaussian process to transfer such information for surrogate modeling. The key novelty of the LOL-GP is a latent regularization model, which identifies regions where transfer should be performed and regions where it should be avoided. Such a ``local transfer'' property is present in many scientific systems: at certain parameters, systems may behave similarly and thus transfer is beneficial; at other parameters, they may behave differently and thus transfer is detrimental. By accounting for local transfer, the LOL-GP can temper the risk of ``negative transfer'', i.e., the risk of worsening predictive performance from information transfer. We derive a Gibbs sampling algorithm for efficient posterior predictive sampling on the LOL-GP, for both the multi-source and multi-fidelity transfer settings. We then show, via a suite of numerical experiments and an application for jet turbine design, the improved surrogate performance of the LOL-GP over existing methods.
♻ ☆ EmissionNet: Air Quality Pollution Forecasting for Agriculture
Air pollution from agricultural emissions is a significant yet often overlooked contributor to environmental and public health challenges. Traditional air quality forecasting models rely on physics-based approaches, which struggle to capture complex, nonlinear pollutant interactions. In this work, we explore forecasting N$_2$O agricultural emissions through evaluating popular architectures, and proposing two novel deep learning architectures, EmissionNet (ENV) and EmissionNet-Transformer (ENT). These models leverage convolutional and transformer-based architectures to extract spatial-temporal dependencies from high-resolution emissions data
comment: The appendix figures are mixed up - several emission plots (e.g. CO2, CH4, GWP) are mislabeled and appear in the wrong order, leading to confusion in interpreting the results
♻ ☆ A Novel Shape-Aware Topological Representation for GPR Data with DNN Integration
Ground Penetrating Radar (GPR) is a widely used Non-Destructive Testing (NDT) technique for subsurface exploration, particularly in infrastructure inspection and maintenance. However, conventional interpretation methods are often limited by noise sensitivity and a lack of structural awareness. This study presents a novel framework that enhances the detection of underground utilities, especially pipelines, by integrating shape-aware topological features derived from B-scan GPR images using Topological Data Analysis (TDA), with the spatial detection capabilities of the YOLOv5 deep neural network (DNN). We propose a novel shape-aware topological representation that amplifies structural features in the input data, thereby improving the model's responsiveness to the geometrical features of buried objects. To address the scarcity of annotated real-world data, we employ a Sim2Real strategy that generates diverse and realistic synthetic datasets, effectively bridging the gap between simulated and real-world domains. Experimental results demonstrate significant improvements in mean Average Precision (mAP), validating the robustness and efficacy of our approach. This approach underscores the potential of TDA-enhanced learning in achieving reliable, real-time subsurface object detection, with broad applications in urban planning, safety inspection, and infrastructure management.
comment: 15 pages, 6 figures
♻ ☆ Task Arithmetic Through The Lens Of One-Shot Federated Learning
Task Arithmetic is a model merging technique that enables the combination of multiple models' capabilities into a single model through simple arithmetic in the weight space, without the need for additional fine-tuning or access to the original training data. However, the factors that determine the success of Task Arithmetic remain unclear. In this paper, we examine Task Arithmetic for multi-task learning by framing it as a one-shot Federated Learning problem. We demonstrate that Task Arithmetic is mathematically equivalent to the commonly used algorithm in Federated Learning, called Federated Averaging (FedAvg). By leveraging well-established theoretical results from FedAvg, we identify two key factors that impact the performance of Task Arithmetic: data heterogeneity and training heterogeneity. To mitigate these challenges, we adapt several algorithms from Federated Learning to improve the effectiveness of Task Arithmetic. Our experiments demonstrate that applying these algorithms can often significantly boost performance of the merged model compared to the original Task Arithmetic approach. This work bridges Task Arithmetic and Federated Learning, offering new theoretical perspectives on Task Arithmetic and improved practical methodologies for model merging.
comment: Published in Transactions on Machine Learning Research
♻ ☆ Navigate the Unknown: Enhancing LLM Reasoning with Intrinsic Motivation Guided Exploration
Reinforcement learning (RL) has emerged as a pivotal method for improving the reasoning capabilities of Large Language Models (LLMs). However, prevalent RL approaches such as Proximal Policy Optimization (PPO) and Group-Regularized Policy Optimization (GRPO) face critical limitations due to their reliance on sparse outcome-based rewards and inadequate mechanisms for incentivizing exploration. These limitations result in inefficient guidance for multi-step reasoning processes. Specifically, sparse reward signals fail to deliver effective or sufficient feedback, particularly for challenging problems. Furthermore, such reward structures induce systematic biases that prioritize exploitation of familiar trajectories over novel solution discovery. These shortcomings critically hinder performance in complex reasoning tasks, which inherently demand iterative refinement across ipntermediate steps. To address these challenges, we propose an Intrinsic Motivation guidEd exploratioN meThOd foR LLM Reasoning (i-MENTOR), a novel method designed to both deliver dense rewards and amplify explorations in the RL-based training paradigm. i-MENTOR introduces three key innovations: trajectory-aware exploration rewards that mitigate bias in token-level strategies while maintaining computational efficiency; dynamic reward scaling to stabilize exploration and exploitation in large action spaces; and advantage-preserving reward implementation that maintains advantage distribution integrity while incorporating exploratory guidance. Experiments across three public datasets demonstrate i-MENTOR's effectiveness with a 22.39% improvement on the difficult dataset Countdown-4.
♻ ☆ Algorithmic contiguity from low-degree conjecture and applications in correlated random graphs
In this paper, assuming a natural strengthening of the low-degree conjecture, we provide evidence of computational hardness for two problems: (1) the (partial) matching recovery problem in the sparse correlated Erd\H{o}s-R\'enyi graphs $\mathcal G(n,q;\rho)$ when the edge-density $q=n^{-1+o(1)}$ and the correlation $\rho<\sqrt{\alpha}$ lies below the Otter's threshold, solving a remaining problem in \cite{DDL23+}; (2) the detection problem between the correlated sparse stochastic block model $\mathcal S(n,\tfrac{\lambda}{n};k,\epsilon;s)$ and a pair of independent stochastic block models $\mathcal S(n,\tfrac{\lambda s}{n};k,\epsilon)$ when $\epsilon^2 \lambda s<1$ lies below the Kesten-Stigum (KS) threshold and $s<\sqrt{\alpha}$ lies below the Otter's threshold, solving a remaining problem in \cite{CDGL24+}. One of the main ingredient in our proof is to derive certain forms of \emph{algorithmic contiguity} between two probability measures based on bounds on their low-degree advantage. To be more precise, consider the high-dimensional hypothesis testing problem between two probability measures $\mathbb{P}$ and $\mathbb{Q}$ based on the sample $\mathsf Y$. We show that if the low-degree advantage $\mathsf{Adv}_{\leq D} \big( \frac{\mathrm{d}\mathbb{P}}{\mathrm{d}\mathbb{Q}} \big)=O(1)$, then (assuming the low-degree conjecture) there is no efficient algorithm $\mathcal A$ such that $\mathbb{Q}(\mathcal A(\mathsf Y)=0)=1-o(1)$ and $\mathbb{P}(\mathcal A(\mathsf Y)=1)=\Omega(1)$. This framework provides a useful tool for performing reductions between different inference tasks.
comment: minor updates; extended abstract to appear in RANDOM 2025
♻ ☆ Thinner Latent Spaces: Detecting Dimension and Imposing Invariance with Conformal Autoencoders
Conformal Autoencoders are a neural network architecture that imposes orthogonality conditions between the gradients of latent variables to obtain disentangled representations of data. In this work we show that orthogonality relations within the latent layer of the network can be leveraged to infer the intrinsic dimensionality of nonlinear manifold data sets (locally characterized by the dimension of their tangent space), while simultaneously computing encoding and decoding (embedding) maps. We outline the relevant theory relying on differential geometry, and describe the corresponding gradient-descent optimization algorithm. The method is applied to several data sets and we highlight its applicability, advantages, and shortcomings. In addition, we demonstrate that the same computational technology can be used to build coordinate invariance to local group actions when defined only on a (reduced) submanifold of the embedding space.
♻ ☆ SpecDec++: Boosting Speculative Decoding via Adaptive Candidate Lengths
Speculative decoding reduces the inference latency of a target large language model via utilizing a smaller and faster draft model. Its performance depends on a hyperparameter K -- the candidate length, i.e., the number of candidate tokens for the target model to verify in each round. However, previous methods often use simple heuristics to choose K, which may result in sub-optimal performance. We study the choice of the candidate length K and formulate it as a Markov Decision Process. We theoretically show that the optimal policy of this Markov decision process takes the form of a threshold policy, i.e., the current speculation should stop and be verified when the probability of getting a rejection exceeds a threshold value. Motivated by this theory, we propose SpecDec++, an enhanced version of speculative decoding that adaptively determines the candidate length on the fly. We augment the draft model with a trained acceptance prediction head to predict the conditional acceptance probability of the candidate tokens. SpecDec++ will stop the current speculation when the predicted probability that at least one token gets rejected exceeds a threshold. We implement SpecDec++ and apply it to the llama-2-chat 7B & 70B model pair. Our adaptive method achieves a 2.04x speedup on the Alpaca dataset (7.2% improvement over the baseline speculative decoding). On the GSM8K and HumanEval datasets, our method achieves a 2.26x speedup (9.4% improvement) and 2.23x speedup (11.1% improvement), respectively. The code of this paper is available at https://github.com/Kaffaljidhmah2/SpecDec_pp.
comment: Accepted to COLM 2025
♻ ☆ PAC-Bayes Analysis for Recalibration in Classification ICML2025
Nonparametric estimation using uniform-width binning is a standard approach for evaluating the calibration performance of machine learning models. However, existing theoretical analyses of the bias induced by binning are limited to binary classification, creating a significant gap with practical applications such as multiclass classification. Additionally, many parametric recalibration algorithms lack theoretical guarantees for their generalization performance. To address these issues, we conduct a generalization analysis of calibration error using the probably approximately correct Bayes framework. This approach enables us to derive the first optimizable upper bound for generalization error in the calibration context. On the basis of our theory, we propose a generalization-aware recalibration algorithm. Numerical experiments show that our algorithm enhances the performance of Gaussian process-based recalibration across various benchmark datasets and models.
comment: Accepted by the 42nd International Conference on Machine Learning (ICML2025), 38 pages, 8 figures
♻ ☆ A Survey on State-of-the-art Deep Learning Applications and Challenges
Deep learning, a branch of artificial intelligence, is a data-driven method that uses multiple layers of interconnected units or neurons to learn intricate patterns and representations directly from raw input data. Empowered by this learning capability, it has become a powerful tool for solving complex problems and is the core driver of many groundbreaking technologies and innovations. Building a deep learning model is challenging due to the algorithm's complexity and the dynamic nature of real-world problems. Several studies have reviewed deep learning concepts and applications. However, the studies mostly focused on the types of deep learning models and convolutional neural network architectures, offering limited coverage of the state-of-the-art deep learning models and their applications in solving complex problems across different domains. Therefore, motivated by the limitations, this study aims to comprehensively review the state-of-the-art deep learning models in computer vision, natural language processing, time series analysis and pervasive computing, and robotics. We highlight the key features of the models and their effectiveness in solving the problems within each domain. Furthermore, this study presents the fundamentals of deep learning, various deep learning model types and prominent convolutional neural network architectures. Finally, challenges and future directions in deep learning research are discussed to offer a broader perspective for future researchers.
comment: Document updated to two-column formatting. This manuscript has been accepted for publication in Engineering Applications of Artificial Intelligence (Elsevier)
♻ ☆ An Outlook on the Opportunities and Challenges of Multi-Agent AI Systems
A multi-agent AI system (MAS) is composed of multiple autonomous agents that interact, exchange information, and make decisions based on internal generative models. Recent advances in large language models and tool-using agents have made MAS increasingly practical in areas like scientific discovery and collaborative automation. However, key questions remain: When are MAS more effective than single-agent systems? What new safety risks arise from agent interactions? And how should we evaluate their reliability and structure? This paper outlines a formal framework for analyzing MAS, focusing on two core aspects: effectiveness and safety. We explore whether MAS truly improve robustness, adaptability, and performance, or merely repackage known techniques like ensemble learning. We also study how inter-agent dynamics may amplify or suppress system vulnerabilities. While MAS are relatively new to the signal processing community, we envision them as a powerful abstraction that extends classical tools like distributed estimation and sensor fusion to higher-level, policy-driven inference. Through experiments on data science automation, we highlight the potential of MAS to reshape how signal processing systems are designed and trusted.
♻ ☆ On the Principles of ReLU Networks with One Hidden Layer
A neural network with one hidden layer or a two-layer network (regardless of the input layer) is the simplest feedforward neural network, whose mechanism may be the basis of more general network architectures. However, even to this type of simple architecture, it is also a ``black box''; that is, it remains unclear how to interpret the mechanism of its solutions obtained by the back-propagation algorithm and how to control the training process through a deterministic way. This paper systematically studies the first problem by constructing universal function-approximation solutions. It is shown that, both theoretically and experimentally, the training solution for the one-dimensional input could be completely understood, and that for a higher-dimensional input can also be well interpreted to some extent. Those results pave the way for thoroughly revealing the black box of two-layer ReLU networks and advance the understanding of deep ReLU networks.
Multimedia 7
☆ Lumos-1: On Autoregressive Video Generation from a Unified Model Perspective
Autoregressive large language models (LLMs) have unified a vast range of language tasks, inspiring preliminary efforts in autoregressive video generation. Existing autoregressive video generators either diverge from standard LLM architectures, depend on bulky external text encoders, or incur prohibitive latency due to next-token decoding. In this paper, we introduce Lumos-1, an autoregressive video generator that retains the LLM architecture with minimal architectural modifications. To inject spatiotemporal correlations in LLMs, we identify the efficacy of incorporating 3D RoPE and diagnose its imbalanced frequency spectrum ranges. Therefore, we propose MM-RoPE, a RoPE scheme that preserves the original textual RoPE while providing comprehensive frequency spectra and scaled 3D positions for modeling multimodal spatiotemporal data. Moreover, Lumos-1 resorts to a token dependency strategy that obeys intra-frame bidirectionality and inter-frame temporal causality. Based on this dependency strategy, we identify the issue of frame-wise loss imbalance caused by spatial information redundancy and solve it by proposing Autoregressive Discrete Diffusion Forcing (AR-DF). AR-DF introduces temporal tube masking during training with a compatible inference-time masking policy to avoid quality degradation. By using memory-efficient training techniques, we pre-train Lumos-1 on only 48 GPUs, achieving performance comparable to EMU3 on GenEval, COSMOS-Video2World on VBench-I2V, and OpenSoraPlan on VBench-T2V. Code and models are available at https://github.com/alibaba-damo-academy/Lumos.
comment: Code and Models: https://github.com/alibaba-damo-academy/Lumos
☆ Visual Semantic Description Generation with MLLMs for Image-Text Matching ICME2025
Image-text matching (ITM) aims to address the fundamental challenge of aligning visual and textual modalities, which inherently differ in their representations, continuous, high-dimensional image features vs. discrete, structured text. We propose a novel framework that bridges the modality gap by leveraging multimodal large language models (MLLMs) as visual semantic parsers. By generating rich Visual Semantic Descriptions (VSD), MLLMs provide semantic anchor that facilitate cross-modal alignment. Our approach combines: (1) Instance-level alignment by fusing visual features with VSD to enhance the linguistic expressiveness of image representations, and (2) Prototype-level alignment through VSD clustering to ensure category-level consistency. These modules can be seamlessly integrated into existing ITM models. Extensive experiments on Flickr30K and MSCOCO demonstrate substantial performance improvements. The approach also exhibits remarkable zero-shot generalization to cross-domain tasks, including news and remote sensing ITM. The code and model checkpoints are available at https://github.com/Image-Text-Matching/VSD.
comment: Accepted by ICME2025 oral
☆ FreeAudio: Training-Free Timing Planning for Controllable Long-Form Text-to-Audio Generation ACM MM 2025
Text-to-audio (T2A) generation has achieved promising results with the recent advances in generative models. However, because of the limited quality and quantity of temporally-aligned audio-text pairs, existing T2A methods struggle to handle the complex text prompts that contain precise timing control, e.g., "owl hooted at 2.4s-5.2s". Recent works have explored data augmentation techniques or introduced timing conditions as model inputs to enable timing-conditioned 10-second T2A generation, while their synthesis quality is still limited. In this work, we propose a novel training-free timing-controlled T2A framework, FreeAudio, making the first attempt to enable timing-controlled long-form T2A generation, e.g., "owl hooted at 2.4s-5.2s and crickets chirping at 0s-24s". Specifically, we first employ an LLM to plan non-overlapping time windows and recaption each with a refined natural language description, based on the input text and timing prompts. Then we introduce: 1) Decoupling and Aggregating Attention Control for precise timing control; 2) Contextual Latent Composition for local smoothness and Reference Guidance for global consistency. Extensive experiments show that: 1) FreeAudio achieves state-of-the-art timing-conditioned T2A synthesis quality among training-free methods and is comparable to leading training-based methods; 2) FreeAudio demonstrates comparable long-form generation quality with training-based Stable Audio and paves the way for timing-controlled long-form T2A synthesis. Demo samples are available at: https://freeaudio.github.io/FreeAudio/
comment: Accepted at ACM MM 2025
☆ PanMatch: Unleashing the Potential of Large Vision Models for Unified Matching Models
This work presents PanMatch, a versatile foundation model for robust correspondence matching. Unlike previous methods that rely on task-specific architectures and domain-specific fine-tuning to support tasks like stereo matching, optical flow or feature matching, our key insight is that any two-frame correspondence matching task can be addressed within a 2D displacement estimation framework using the same model weights. Such a formulation eliminates the need for designing specialized unified architectures or task-specific ensemble models. Instead, it achieves multi-task integration by endowing displacement estimation algorithms with unprecedented generalization capabilities. To this end, we highlight the importance of a robust feature extractor applicable across multiple domains and tasks, and propose the feature transformation pipeline that leverage all-purpose features from Large Vision Models to endow matching baselines with zero-shot cross-view matching capabilities. Furthermore, we assemble a cross-domain dataset with near 1.8 million samples from stereo matching, optical flow, and feature matching domains to pretrain PanMatch. We demonstrate the versatility of PanMatch across a wide range of domains and downstream tasks using the same model weights. Our model outperforms UniMatch and Flow-Anything on cross-task evaluations, and achieves comparable performance to most state-of-the-art task-specific algorithms on task-oriented benchmarks. Additionally, PanMatch presents unprecedented zero-shot performance in abnormal scenarios, such as rainy day and satellite imagery, where most existing robust algorithms fail to yield meaningful results.
☆ Infinite Video Understanding
The rapid advancements in Large Language Models (LLMs) and their multimodal extensions (MLLMs) have ushered in remarkable progress in video understanding. However, a fundamental challenge persists: effectively processing and comprehending video content that extends beyond minutes or hours. While recent efforts like Video-XL-2 have demonstrated novel architectural solutions for extreme efficiency, and advancements in positional encoding such as HoPE and VideoRoPE++ aim to improve spatio-temporal understanding over extensive contexts, current state-of-the-art models still encounter significant computational and memory constraints when faced with the sheer volume of visual tokens from lengthy sequences. Furthermore, maintaining temporal coherence, tracking complex events, and preserving fine-grained details over extended periods remain formidable hurdles, despite progress in agentic reasoning systems like Deep Video Discovery. This position paper posits that a logical, albeit ambitious, next frontier for multimedia research is Infinite Video Understanding -- the capability for models to continuously process, understand, and reason about video data of arbitrary, potentially never-ending duration. We argue that framing Infinite Video Understanding as a blue-sky research objective provides a vital north star for the multimedia, and the wider AI, research communities, driving innovation in areas such as streaming architectures, persistent memory mechanisms, hierarchical and adaptive representations, event-centric reasoning, and novel evaluation paradigms. Drawing inspiration from recent work on long/ultra-long video understanding and several closely related fields, we outline the core challenges and key research directions towards achieving this transformative capability.
♻ ☆ Token Communications: A Unified Framework for Cross-modal Context-aware Semantic Communications
In this paper, we introduce token communications (TokCom), a large model-driven framework to leverage cross-modal context information in generative semantic communications (GenSC). TokCom is a new paradigm, motivated by the recent success of generative foundation models and multimodal large language models (GFM/MLLMs), where the communication units are tokens, enabling efficient transformer-based token processing at the transmitter and receiver. In this paper, we introduce the potential opportunities and challenges of leveraging context in GenSC, explore how to integrate GFM/MLLMs-based token processing into semantic communication systems to leverage cross-modal context effectively at affordable complexity, present the key principles for efficient TokCom at various layers in future wireless networks. In a typical image semantic communication setup, we demonstrate a significant improvement of the bandwidth efficiency, achieved by TokCom by leveraging the context information among tokens. Finally, the potential research directions are identified to facilitate adoption of TokCom in future wireless networks.
comment: Accepted at IEEE Wireless Communications Magazine
♻ ☆ T-GVC: Trajectory-Guided Generative Video Coding at Ultra-Low Bitrates
Recent advances in video generation techniques have given rise to an emerging paradigm of generative video coding, aiming to achieve semantically accurate reconstructions in Ultra-Low Bitrate (ULB) scenarios by leveraging strong generative priors. However, most existing methods are limited by domain specificity (e.g., facial or human videos) or an excessive dependence on high-level text guidance, which often fails to capture motion details and results in unrealistic reconstructions. To address these challenges, we propose a Trajectory-Guided Generative Video Coding framework (dubbed T-GVC). T-GVC employs a semantic-aware sparse motion sampling pipeline to effectively bridge low-level motion tracking with high-level semantic understanding by extracting pixel-wise motion as sparse trajectory points based on their semantic importance, not only significantly reducing the bitrate but also preserving critical temporal semantic information. In addition, by incorporating trajectory-aligned loss constraints into diffusion processes, we introduce a training-free latent space guidance mechanism to ensure physically plausible motion patterns without sacrificing the inherent capabilities of generative models. Experimental results demonstrate that our framework outperforms both traditional codecs and state-of-the-art end-to-end video compression methods under ULB conditions. Furthermore, additional experiments confirm that our approach achieves more precise motion control than existing text-guided methods, paving the way for a novel direction of generative video coding guided by geometric motion modeling.
Computation and Language 112
☆ Traceable Evidence Enhanced Visual Grounded Reasoning: Evaluation and Methodology
Models like OpenAI-o3 pioneer visual grounded reasoning by dynamically referencing visual regions, just like human "thinking with images". However, no benchmark exists to evaluate these capabilities holistically. To bridge this gap, we propose TreeBench (Traceable Evidence Evaluation Benchmark), a diagnostic benchmark built on three principles: (1) focused visual perception of subtle targets in complex scenes, (2) traceable evidence via bounding box evaluation, and (3) second-order reasoning to test object interactions and spatial hierarchies beyond simple object localization. Prioritizing images with dense objects, we initially sample 1K high-quality images from SA-1B, and incorporate eight LMM experts to manually annotate questions, candidate options, and answers for each image. After three stages of quality control, TreeBench consists of 405 challenging visual question-answering pairs, even the most advanced models struggle with this benchmark, where none of them reach 60% accuracy, e.g., OpenAI-o3 scores only 54.87. Furthermore, we introduce TreeVGR (Traceable Evidence Enhanced Visual Grounded Reasoning), a training paradigm to supervise localization and reasoning jointly with reinforcement learning, enabling accurate localizations and explainable reasoning pathways. Initialized from Qwen2.5-VL-7B, it improves V* Bench (+16.8), MME-RealWorld (+12.6), and TreeBench (+13.4), proving traceability is key to advancing vision-grounded reasoning. The code is available at https://github.com/Haochen-Wang409/TreeVGR.
☆ PyVision: Agentic Vision with Dynamic Tooling
LLMs are increasingly deployed as agents, systems capable of planning, reasoning, and dynamically calling external tools. However, in visual reasoning, prior approaches largely remain limited by predefined workflows and static toolsets. In this report, we present PyVision, an interactive, multi-turn framework that enables MLLMs to autonomously generate, execute, and refine Python-based tools tailored to the task at hand, unlocking flexible and interpretable problem-solving. We develop a taxonomy of the tools created by PyVision and analyze their usage across a diverse set of benchmarks. Quantitatively, PyVision achieves consistent performance gains, boosting GPT-4.1 by +7.8% on V* and Claude-4.0-Sonnet by +31.1% on VLMsAreBlind-mini. These results point to a broader shift: dynamic tooling allows models not just to use tools, but to invent them, advancing toward more agentic visual reasoning.
comment: 26 Pages, 10 Figures, Technical report
☆ Automating Expert-Level Medical Reasoning Evaluation of Large Language Models
As large language models (LLMs) become increasingly integrated into clinical decision-making, ensuring transparent and trustworthy reasoning is essential. However, existing evaluation strategies of LLMs' medical reasoning capability either suffer from unsatisfactory assessment or poor scalability, and a rigorous benchmark remains lacking. To address this, we introduce MedThink-Bench, a benchmark designed for rigorous, explainable, and scalable assessment of LLMs' medical reasoning. MedThink-Bench comprises 500 challenging questions across ten medical domains, each annotated with expert-crafted step-by-step rationales. Building on this, we propose LLM-w-Ref, a novel evaluation framework that leverages fine-grained rationales and LLM-as-a-Judge mechanisms to assess intermediate reasoning with expert-level fidelity while maintaining scalability. Experiments show that LLM-w-Ref exhibits a strong positive correlation with expert judgments. Benchmarking twelve state-of-the-art LLMs, we find that smaller models (e.g., MedGemma-27B) can surpass larger proprietary counterparts (e.g., OpenAI-o3). Overall, MedThink-Bench offers a foundational tool for evaluating LLMs' medical reasoning, advancing their safe and responsible deployment in clinical practice.
comment: 22 pages,6 figures
☆ Performance and Practical Considerations of Large and Small Language Models in Clinical Decision Support in Rheumatology
Large language models (LLMs) show promise for supporting clinical decision-making in complex fields such as rheumatology. Our evaluation shows that smaller language models (SLMs), combined with retrieval-augmented generation (RAG), achieve higher diagnostic and therapeutic performance than larger models, while requiring substantially less energy and enabling cost-efficient, local deployment. These features are attractive for resource-limited healthcare. However, expert oversight remains essential, as no model consistently reached specialist-level accuracy in rheumatology.
☆ Why is Your Language Model a Poor Implicit Reward Model?
Reward models are key to language model post-training and inference pipelines. Conveniently, recent work showed that every language model defines an implicit reward model (IM-RM), without requiring any architectural changes. However, such IM-RMs tend to generalize worse, especially out-of-distribution, compared to explicit reward models (EX-RMs) that apply a dedicated linear head over the hidden representations of a language model. The existence of a generalization gap is puzzling, as EX-RMs and IM-RMs are nearly identical. They can be trained using the same data, loss function, and language model, and differ only in how the reward is computed. Towards a fundamental understanding of the implicit biases underlying different reward model types, we investigate the root cause of this gap. Our main finding, backed by theory and experiments, is that IM-RMs rely more heavily on superficial token-level cues. Consequently, they often generalize worse than EX-RMs under token-level distribution shifts, as well as in-distribution. Furthermore, we provide evidence against alternative hypotheses for the generalization gap. Most notably, we challenge the intuitive claim that IM-RMs struggle in tasks where generation is harder than verification because they can operate both as a verifier and a generator. Taken together, our results highlight that seemingly minor design choices can substantially impact the generalization behavior of reward models.
☆ Scaling RL to Long Videos
We introduce a full-stack framework that scales up reasoning in vision-language models (VLMs) to long videos, leveraging reinforcement learning. We address the unique challenges of long video reasoning by integrating three critical components: (1) a large-scale dataset, LongVideo-Reason, comprising 52K long video QA pairs with high-quality reasoning annotations across diverse domains such as sports, games, and vlogs; (2) a two-stage training pipeline that extends VLMs with chain-of-thought supervised fine-tuning (CoT-SFT) and reinforcement learning (RL); and (3) a training infrastructure for long video RL, named Multi-modal Reinforcement Sequence Parallelism (MR-SP), which incorporates sequence parallelism and a vLLM-based engine tailored for long video, using cached video embeddings for efficient rollout and prefilling. In experiments, LongVILA-R1-7B achieves strong performance on long video QA benchmarks such as VideoMME. It also outperforms Video-R1-7B and even matches Gemini-1.5-Pro across temporal reasoning, goal and purpose reasoning, spatial reasoning, and plot reasoning on our LongVideo-Reason-eval benchmark. Notably, our MR-SP system achieves up to 2.1x speedup on long video RL training. LongVILA-R1 demonstrates consistent performance gains as the number of input video frames scales. LongVILA-R1 marks a firm step towards long video reasoning in VLMs. In addition, we release our training system for public availability that supports RL training on various modalities (video, text, and audio), various models (VILA and Qwen series), and even image and video generation models. On a single A100 node (8 GPUs), it supports RL training on hour-long videos (e.g., 3,600 frames / around 256k tokens).
comment: Code and models are available at https://github.com/NVlabs/Long-RL
☆ MIRIX: Multi-Agent Memory System for LLM-Based Agents
Although memory capabilities of AI agents are gaining increasing attention, existing solutions remain fundamentally limited. Most rely on flat, narrowly scoped memory components, constraining their ability to personalize, abstract, and reliably recall user-specific information over time. To this end, we introduce MIRIX, a modular, multi-agent memory system that redefines the future of AI memory by solving the field's most critical challenge: enabling language models to truly remember. Unlike prior approaches, MIRIX transcends text to embrace rich visual and multimodal experiences, making memory genuinely useful in real-world scenarios. MIRIX consists of six distinct, carefully structured memory types: Core, Episodic, Semantic, Procedural, Resource Memory, and Knowledge Vault, coupled with a multi-agent framework that dynamically controls and coordinates updates and retrieval. This design enables agents to persist, reason over, and accurately retrieve diverse, long-term user data at scale. We validate MIRIX in two demanding settings. First, on ScreenshotVQA, a challenging multimodal benchmark comprising nearly 20,000 high-resolution computer screenshots per sequence, requiring deep contextual understanding and where no existing memory systems can be applied, MIRIX achieves 35% higher accuracy than the RAG baseline while reducing storage requirements by 99.9%. Second, on LOCOMO, a long-form conversation benchmark with single-modal textual input, MIRIX attains state-of-the-art performance of 85.4%, far surpassing existing baselines. These results show that MIRIX sets a new performance standard for memory-augmented LLM agents. To allow users to experience our memory system, we provide a packaged application powered by MIRIX. It monitors the screen in real time, builds a personalized memory base, and offers intuitive visualization and secure local storage to ensure privacy.
☆ SAGE: A Visual Language Model for Anomaly Detection via Fact Enhancement and Entropy-aware Alignment
While Vision-Language Models (VLMs) have shown promising progress in general multimodal tasks, they often struggle in industrial anomaly detection and reasoning, particularly in delivering interpretable explanations and generalizing to unseen categories. This limitation stems from the inherently domain-specific nature of anomaly detection, which hinders the applicability of existing VLMs in industrial scenarios that require precise, structured, and context-aware analysis. To address these challenges, we propose SAGE, a VLM-based framework that enhances anomaly reasoning through Self-Guided Fact Enhancement (SFE) and Entropy-aware Direct Preference Optimization (E-DPO). SFE integrates domain-specific knowledge into visual reasoning via fact extraction and fusion, while E-DPO aligns model outputs with expert preferences using entropy-aware optimization. Additionally, we introduce AD-PL, a preference-optimized dataset tailored for industrial anomaly reasoning, consisting of 28,415 question-answering instances with expert-ranked responses. To evaluate anomaly reasoning models, we develop Multiscale Logical Evaluation (MLE), a quantitative framework analyzing model logic and consistency. SAGE demonstrates superior performance on industrial anomaly datasets under zero-shot and one-shot settings. The code, model and dataset are available at https://github.com/amoreZgx1n/SAGE.
comment: Accepted by ACMMM2025
☆ DTECT: Dynamic Topic Explorer & Context Tracker
The explosive growth of textual data over time presents a significant challenge in uncovering evolving themes and trends. Existing dynamic topic modeling techniques, while powerful, often exist in fragmented pipelines that lack robust support for interpretation and user-friendly exploration. We introduce DTECT (Dynamic Topic Explorer & Context Tracker), an end-to-end system that bridges the gap between raw textual data and meaningful temporal insights. DTECT provides a unified workflow that supports data preprocessing, multiple model architectures, and dedicated evaluation metrics to analyze the topic quality of temporal topic models. It significantly enhances interpretability by introducing LLM-driven automatic topic labeling, trend analysis via temporally salient words, interactive visualizations with document-level summarization, and a natural language chat interface for intuitive data querying. By integrating these features into a single, cohesive platform, DTECT empowers users to more effectively track and understand thematic dynamics. DTECT is open-source and available at https://github.com/AdhyaSuman/DTECT.
comment: Code: https://github.com/AdhyaSuman/DTECT | Demo: https://huggingface.co/spaces/AdhyaSuman/DTECT | Video: https://youtu.be/B8nNfxFoJAU
☆ Automating MD simulations for Proteins using Large language Models: NAMD-Agent
Molecular dynamics simulations are an essential tool in understanding protein structure, dynamics, and function at the atomic level. However, preparing high quality input files for MD simulations can be a time consuming and error prone process. In this work, we introduce an automated pipeline that leverages Large Language Models (LLMs), specifically Gemini 2.0 Flash, in conjunction with python scripting and Selenium based web automation to streamline the generation of MD input files. The pipeline exploits CHARMM GUI's comprehensive web-based interface for preparing simulation-ready inputs for NAMD. By integrating Gemini's code generation and iterative refinement capabilities, simulation scripts are automatically written, executed, and revised to navigate CHARMM GUI, extract appropriate parameters, and produce the required NAMD input files. Post processing is performed using additional software to further refine the simulation outputs, thereby enabling a complete and largely hands free workflow. Our results demonstrate that this approach reduces setup time, minimizes manual errors, and offers a scalable solution for handling multiple protein systems in parallel. This automated framework paves the way for broader application of LLMs in computational structural biology, offering a robust and adaptable platform for future developments in simulation automation.
comment: 34 pages
☆ DocCHA: Towards LLM-Augmented Interactive Online diagnosis System
Despite the impressive capabilities of Large Language Models (LLMs), existing Conversational Health Agents (CHAs) remain static and brittle, incapable of adaptive multi-turn reasoning, symptom clarification, or transparent decision-making. This hinders their real-world applicability in clinical diagnosis, where iterative and structured dialogue is essential. We propose DocCHA, a confidence-aware, modular framework that emulates clinical reasoning by decomposing the diagnostic process into three stages: (1) symptom elicitation, (2) history acquisition, and (3) causal graph construction. Each module uses interpretable confidence scores to guide adaptive questioning, prioritize informative clarifications, and refine weak reasoning links. Evaluated on two real-world Chinese consultation datasets (IMCS21, DX), DocCHA consistently outperforms strong prompting-based LLM baselines (GPT-3.5, GPT-4o, LLaMA-3), achieving up to 5.18 percent higher diagnostic accuracy and over 30 percent improvement in symptom recall, with only modest increase in dialogue turns. These results demonstrate the effectiveness of DocCHA in enabling structured, transparent, and efficient diagnostic conversations -- paving the way for trustworthy LLM-powered clinical assistants in multilingual and resource-constrained settings.
☆ Alpay Algebra V: Multi-Layered Semantic Games and Transfinite Fixed-Point Simulation
This paper extends the self-referential framework of Alpay Algebra into a multi-layered semantic game architecture where transfinite fixed-point convergence encompasses hierarchical sub-games at each iteration level. Building upon Alpay Algebra IV's empathetic embedding concept, we introduce a nested game-theoretic structure where the alignment process between AI systems and documents becomes a meta-game containing embedded decision problems. We formalize this through a composite operator $\phi(\cdot, \gamma(\cdot))$ where $\phi$ drives the main semantic convergence while $\gamma$ resolves local sub-games. The resulting framework demonstrates that game-theoretic reasoning emerges naturally from fixed-point iteration rather than being imposed externally. We prove a Game Theorem establishing existence and uniqueness of semantic equilibria under realistic cognitive simulation assumptions. Our verification suite includes adaptations of Banach's fixed-point theorem to transfinite contexts, a novel $\phi$-topology based on the Kozlov-Maz'ya-Rossmann formula for handling semantic singularities, and categorical consistency tests via the Yoneda lemma. The paper itself functions as a semantic artifact designed to propagate its fixed-point patterns in AI embedding spaces -- a deliberate instantiation of the "semantic virus" concept it theorizes. All results are grounded in category theory, information theory, and realistic AI cognition models, ensuring practical applicability beyond pure mathematical abstraction.
comment: 18 pages, 2 figures
☆ From Ambiguity to Accuracy: The Transformative Effect of Coreference Resolution on Retrieval-Augmented Generation systems
Retrieval-Augmented Generation (RAG) has emerged as a crucial framework in natural language processing (NLP), improving factual consistency and reducing hallucinations by integrating external document retrieval with large language models (LLMs). However, the effectiveness of RAG is often hindered by coreferential complexity in retrieved documents, introducing ambiguity that disrupts in-context learning. In this study, we systematically investigate how entity coreference affects both document retrieval and generative performance in RAG-based systems, focusing on retrieval relevance, contextual understanding, and overall response quality. We demonstrate that coreference resolution enhances retrieval effectiveness and improves question-answering (QA) performance. Through comparative analysis of different pooling strategies in retrieval tasks, we find that mean pooling demonstrates superior context capturing ability after applying coreference resolution. In QA tasks, we discover that smaller models benefit more from the disambiguation process, likely due to their limited inherent capacity for handling referential ambiguity. With these findings, this study aims to provide a deeper understanding of the challenges posed by coreferential complexity in RAG, providing guidance for improving retrieval and generation in knowledge-intensive AI applications.
☆ Conditional Unigram Tokenization with Parallel Data ICML 2025
We introduce conditional unigram tokenization, a novel approach that extends unigram tokenization by conditioning target token probabilities on source-language tokens from parallel data. Given a fixed source tokenizer, our method learns a target tokenizer that maximizes cross-lingual semantic alignment. We evaluate our tokenizer on four language pairs across different families and resource levels, examining intrinsic properties and downstream performance on machine translation and language modeling. While our conditional tokenizer maintains comparable statistical properties to standard unigram tokenizers, results are mixed: we observe no improvements in machine translation quality, but find consistent perplexity reductions in language modeling. We hypothesize that quadratic scaling of conditional probability estimation with respect to the vocabulary size creates a data efficiency bottleneck. Our findings suggest that alternative parameterizations may be necessary for practical cross-lingual tokenization.
comment: 21 pages, 4 figures, submitted to Tokenization Workshop (TokShop) at ICML 2025
☆ On the Effect of Instruction Tuning Loss on Generalization ACL
Instruction Tuning has emerged as a pivotal post-training paradigm that enables pre-trained language models to better follow user instructions. Despite its significance, little attention has been given to optimizing the loss function used. A fundamental, yet often overlooked, question is whether the conventional auto-regressive objective - where loss is computed only on response tokens, excluding prompt tokens - is truly optimal for instruction tuning. In this work, we systematically investigate the impact of differentially weighting prompt and response tokens in instruction tuning loss, and propose Weighted Instruction Tuning (WIT) as a better alternative to conventional instruction tuning. Through extensive experiments on five language models of different families and scale, three finetuning datasets of different sizes, and five diverse evaluation benchmarks, we show that the standard instruction tuning loss often yields suboptimal performance and limited robustness to input prompt variations. We find that a low-to-moderate weight for prompt tokens coupled with a moderate-to-high weight for response tokens yields the best-performing models across settings and also serve as better starting points for the subsequent preference alignment training. These findings highlight the need to reconsider instruction tuning loss and offer actionable insights for developing more robust and generalizable models. Our code is open-sourced at https://github.com/kowndinya-renduchintala/WIT.
comment: Transactions of the Association for Computational Linguistics (TACL)
☆ Understanding and Controlling Repetition Neurons and Induction Heads in In-Context Learning
This paper investigates the relationship between large language models' (LLMs) ability to recognize repetitive input patterns and their performance on in-context learning (ICL). In contrast to prior work that has primarily focused on attention heads, we examine this relationship from the perspective of skill neurons, specifically repetition neurons. Our experiments reveal that the impact of these neurons on ICL performance varies depending on the depth of the layer in which they reside. By comparing the effects of repetition neurons and induction heads, we further identify strategies for reducing repetitive outputs while maintaining strong ICL capabilities.
☆ Bridging Logic and Learning: Decoding Temporal Logic Embeddings via Transformers ECML-PKDD
Continuous representations of logic formulae allow us to integrate symbolic knowledge into data-driven learning algorithms. If such embeddings are semantically consistent, i.e. if similar specifications are mapped into nearby vectors, they enable continuous learning and optimization directly in the semantic space of formulae. However, to translate the optimal continuous representation into a concrete requirement, such embeddings must be invertible. We tackle this issue by training a Transformer-based decoder-only model to invert semantic embeddings of Signal Temporal Logic (STL) formulae. STL is a powerful formalism that allows us to describe properties of signals varying over time in an expressive yet concise way. By constructing a small vocabulary from STL syntax, we demonstrate that our proposed model is able to generate valid formulae after only 1 epoch and to generalize to the semantics of the logic in about 10 epochs. Additionally, the model is able to decode a given embedding into formulae that are often simpler in terms of length and nesting while remaining semantically close (or equivalent) to gold references. We show the effectiveness of our methodology across various levels of training formulae complexity to assess the impact of training data on the model's ability to effectively capture the semantic information contained in the embeddings and generalize out-of-distribution. Finally, we deploy our model for solving a requirement mining task, i.e. inferring STL specifications that solve a classification task on trajectories, performing the optimization directly in the semantic space.
comment: 16 pages, 3 figures, to be published in ECML-PKDD
☆ StreamUni: Achieving Streaming Speech Translation with a Unified Large Speech-Language Model
Streaming speech translation (StreamST) requires determining appropriate timing, known as policy, to generate translations while continuously receiving source speech inputs, balancing low latency with high translation quality. However, existing StreamST methods typically operate on sentence-level speech segments, referred to as simultaneous speech translation (SimulST). In practice, they require collaboration with segmentation models to accomplish StreamST, where the truncated speech segments constrain SimulST models to make policy decisions and generate translations based on limited contextual information. Moreover, SimulST models struggle to learn effective policies due to the complexity of speech inputs and cross-lingual generation. To address these challenges, we propose StreamUni, which achieves StreamST through a unified Large Speech-Language Model (LSLM). Specifically, StreamUni incorporates speech Chain-of-Thought (CoT) in guiding the LSLM to generate multi-stage outputs. Leveraging these multi-stage outputs, StreamUni simultaneously accomplishes speech segmentation, policy decision, and translation generation, completing StreamST without requiring massive policy-specific training. Additionally, we propose a streaming CoT training method that enhances low-latency policy decisions and generation capabilities using limited CoT data. Experiments demonstrate that our approach achieves state-of-the-art performance on StreamST tasks.
comment: The code is at https://github.com/ictnlp/StreamUni; The model is at https://huggingface.co/ICTNLP/StreamUni-Phi4
☆ When Large Language Models Meet Law: Dual-Lens Taxonomy, Technical Advances, and Ethical Governance
This paper establishes the first comprehensive review of Large Language Models (LLMs) applied within the legal domain. It pioneers an innovative dual lens taxonomy that integrates legal reasoning frameworks and professional ontologies to systematically unify historical research and contemporary breakthroughs. Transformer-based LLMs, which exhibit emergent capabilities such as contextual reasoning and generative argumentation, surmount traditional limitations by dynamically capturing legal semantics and unifying evidence reasoning. Significant progress is documented in task generalization, reasoning formalization, workflow integration, and addressing core challenges in text processing, knowledge integration, and evaluation rigor via technical innovations like sparse attention mechanisms and mixture-of-experts architectures. However, widespread adoption of LLM introduces critical challenges: hallucination, explainability deficits, jurisdictional adaptation difficulties, and ethical asymmetry. This review proposes a novel taxonomy that maps legal roles to NLP subtasks and computationally implements the Toulmin argumentation framework, thus systematizing advances in reasoning, retrieval, prediction, and dispute resolution. It identifies key frontiers including low-resource systems, multimodal evidence integration, and dynamic rebuttal handling. Ultimately, this work provides both a technical roadmap for researchers and a conceptual framework for practitioners navigating the algorithmic future, laying a robust foundation for the next era of legal artificial intelligence. We have created a GitHub repository to index the relevant papers: https://github.com/Kilimajaro/LLMs_Meet_Law.
☆ Code-Switching in End-to-End Automatic Speech Recognition: A Systematic Literature Review
Motivated by a growing research interest into automatic speech recognition (ASR), and the growing body of work for languages in which code-switching (CS) often occurs, we present a systematic literature review of code-switching in end-to-end ASR models. We collect and manually annotate papers published in peer reviewed venues. We document the languages considered, datasets, metrics, model choices, and performance, and present a discussion of challenges in end-to-end ASR for code-switching. Our analysis thus provides insights on current research efforts and available resources as well as opportunities and gaps to guide future research.
☆ GuardVal: Dynamic Large Language Model Jailbreak Evaluation for Comprehensive Safety Testing
Jailbreak attacks reveal critical vulnerabilities in Large Language Models (LLMs) by causing them to generate harmful or unethical content. Evaluating these threats is particularly challenging due to the evolving nature of LLMs and the sophistication required in effectively probing their vulnerabilities. Current benchmarks and evaluation methods struggle to fully address these challenges, leaving gaps in the assessment of LLM vulnerabilities. In this paper, we review existing jailbreak evaluation practices and identify three assumed desiderata for an effective jailbreak evaluation protocol. To address these challenges, we introduce GuardVal, a new evaluation protocol that dynamically generates and refines jailbreak prompts based on the defender LLM's state, providing a more accurate assessment of defender LLMs' capacity to handle safety-critical situations. Moreover, we propose a new optimization method that prevents stagnation during prompt refinement, ensuring the generation of increasingly effective jailbreak prompts that expose deeper weaknesses in the defender LLMs. We apply this protocol to a diverse set of models, from Mistral-7b to GPT-4, across 10 safety domains. Our findings highlight distinct behavioral patterns among the models, offering a comprehensive view of their robustness. Furthermore, our evaluation process deepens the understanding of LLM behavior, leading to insights that can inform future research and drive the development of more secure models.
comment: 24 pages
☆ Not All Preferences are What You Need for Post-Training: Selective Alignment Strategy for Preference Optimization
Post-training alignment of large language models (LLMs) is a critical challenge, as not all tokens contribute equally to model performance. This paper introduces a selective alignment strategy that prioritizes high-impact tokens within preference pairs, leveraging token-level log-probability differences between the current policy and a reference model. By focusing on these informative tokens, our approach reduces computational overhead and enhances alignment fidelity. We further explore the role of reference model quality, demonstrating that stronger reference models significantly improve token selection accuracy and overall optimization effectiveness. Comprehensive experiments on benchmarks such as Arena-Hard and MT-Bench validate the superiority of our Selective-DPO method over standard DPO and distillation-based baselines. Our findings highlight the importance of token-level optimization and reference model selection in advancing preference alignment for LLMs. The code is available at https://github.com/Dongzhijin/SDPO.
☆ Rethinking the Privacy of Text Embeddings: A Reproducibility Study of "Text Embeddings Reveal (Almost) As Much As Text" RecSys 2025
Text embeddings are fundamental to many natural language processing (NLP) tasks, extensively applied in domains such as recommendation systems and information retrieval (IR). Traditionally, transmitting embeddings instead of raw text has been seen as privacy-preserving. However, recent methods such as Vec2Text challenge this assumption by demonstrating that controlled decoding can successfully reconstruct original texts from black-box embeddings. The unexpectedly strong results reported by Vec2Text motivated us to conduct further verification, particularly considering the typically non-intuitive and opaque structure of high-dimensional embedding spaces. In this work, we reproduce the Vec2Text framework and evaluate it from two perspectives: (1) validating the original claims, and (2) extending the study through targeted experiments. First, we successfully replicate the original key results in both in-domain and out-of-domain settings, with only minor discrepancies arising due to missing artifacts, such as model checkpoints and dataset splits. Furthermore, we extend the study by conducting a parameter sensitivity analysis, evaluating the feasibility of reconstructing sensitive inputs (e.g., passwords), and exploring embedding quantization as a lightweight privacy defense. Our results show that Vec2Text is effective under ideal conditions, capable of reconstructing even password-like sequences that lack clear semantics. However, we identify key limitations, including its sensitivity to input sequence length. We also find that Gaussian noise and quantization techniques can mitigate the privacy risks posed by Vec2Text, with quantization offering a simpler and more widely applicable solution. Our findings emphasize the need for caution in using text embeddings and highlight the importance of further research into robust defense mechanisms for NLP systems.
comment: This paper has been accepted for oral presentation in the reproducibility track at RecSys 2025
☆ KeyKnowledgeRAG (K^2RAG): An Enhanced RAG method for improved LLM question-answering capabilities
Fine-tuning is an immensely resource-intensive process when retraining Large Language Models (LLMs) to incorporate a larger body of knowledge. Although many fine-tuning techniques have been developed to reduce the time and computational cost involved, the challenge persists as LLMs continue to grow in size and complexity. To address this, a new approach to knowledge expansion in LLMs is needed. Retrieval-Augmented Generation (RAG) offers one such alternative by storing external knowledge in a database and retrieving relevant chunks to support question answering. However, naive implementations of RAG face significant limitations in scalability and answer accuracy. This paper introduces KeyKnowledgeRAG (K2RAG), a novel framework designed to overcome these limitations. Inspired by the divide-and-conquer paradigm, K2RAG integrates dense and sparse vector search, knowledge graphs, and text summarization to improve retrieval quality and system efficiency. The framework also includes a preprocessing step that summarizes the training data, significantly reducing the training time. K2RAG was evaluated using the MultiHopRAG dataset, where the proposed pipeline was trained on the document corpus and tested on a separate evaluation set. Results demonstrated notable improvements over common naive RAG implementations. K2RAG achieved the highest mean answer similarity score of 0.57, and reached the highest third quartile (Q3) similarity of 0.82, indicating better alignment with ground-truth answers. In addition to improved accuracy, the framework proved highly efficient. The summarization step reduced the average training time of individual components by 93%, and execution speed was up to 40% faster than traditional knowledge graph-based RAG systems. K2RAG also demonstrated superior scalability, requiring three times less VRAM than several naive RAG implementations tested in this study.
comment: 21 pages, 14 figures
☆ SAS: Simulated Attention Score
The attention mechanism is a core component of the Transformer architecture. Various methods have been developed to compute attention scores, including multi-head attention (MHA), multi-query attention, group-query attention and so on. We further analyze the MHA and observe that its performance improves as the number of attention heads increases, provided the hidden size per head remains sufficiently large. Therefore, increasing both the head count and hidden size per head with minimal parameter overhead can lead to significant performance gains at a low cost. Motivated by this insight, we introduce Simulated Attention Score (SAS), which maintains a compact model size while simulating a larger number of attention heads and hidden feature dimension per head. This is achieved by projecting a low-dimensional head representation into a higher-dimensional space, effectively increasing attention capacity without increasing parameter count. Beyond the head representations, we further extend the simulation approach to feature dimension of the key and query embeddings, enhancing expressiveness by mimicking the behavior of a larger model while preserving the original model size. To control the parameter cost, we also propose Parameter-Efficient Attention Aggregation (PEAA). Comprehensive experiments on a variety of datasets and tasks demonstrate the effectiveness of the proposed SAS method, achieving significant improvements over different attention variants.
comment: Tech Report
☆ An Automated Length-Aware Quality Metric for Summarization
This paper proposes NOrmed Index of Retention (NOIR), a quantitative objective metric for evaluating summarization quality of arbitrary texts that relies on both the retention of semantic meaning and the summary length compression. This gives a measure of how well the recall-compression tradeoff is managed, the most important skill in summarization. Experiments demonstrate that NOIR effectively captures the token-length / semantic retention tradeoff of a summarizer and correlates to human perception of sumarization quality. Using a language model-embedding to measure semantic similarity, it provides an automated alternative for assessing summarization quality without relying on time-consuming human-generated reference summaries. The proposed metric can be applied to various summarization tasks, offering an automated tool for evaluating and improving summarization algorithms, summarization prompts, and synthetically-generated summaries.
☆ Lost in Pronunciation: Detecting Chinese Offensive Language Disguised by Phonetic Cloaking Replacement
Phonetic Cloaking Replacement (PCR), defined as the deliberate use of homophonic or near-homophonic variants to hide toxic intent, has become a major obstacle to Chinese content moderation. While this problem is well-recognized, existing evaluations predominantly rely on rule-based, synthetic perturbations that ignore the creativity of real users. We organize PCR into a four-way surface-form taxonomy and compile \ours, a dataset of 500 naturally occurring, phonetically cloaked offensive posts gathered from the RedNote platform. Benchmarking state-of-the-art LLMs on this dataset exposes a serious weakness: the best model reaches only an F1-score of 0.672, and zero-shot chain-of-thought prompting pushes performance even lower. Guided by error analysis, we revisit a Pinyin-based prompting strategy that earlier studies judged ineffective and show that it recovers much of the lost accuracy. This study offers the first comprehensive taxonomy of Chinese PCR, a realistic benchmark that reveals current detectors' limits, and a lightweight mitigation technique that advances research on robust toxicity detection.
comment: In progress
☆ FrugalRAG: Learning to retrieve and reason for multi-hop QA ICML
We consider the problem of answering complex questions, given access to a large unstructured document corpus. The de facto approach to solving the problem is to leverage language models that (iteratively) retrieve and reason through the retrieved documents, until the model has sufficient information to generate an answer. Attempts at improving this approach focus on retrieval-augmented generation (RAG) metrics such as accuracy and recall and can be categorized into two types: (a) fine-tuning on large question answering (QA) datasets augmented with chain-of-thought traces, and (b) leveraging RL-based fine-tuning techniques that rely on question-document relevance signals. However, efficiency in the number of retrieval searches is an equally important metric, which has received less attention. In this work, we show that: (1) Large-scale fine-tuning is not needed to improve RAG metrics, contrary to popular claims in recent literature. Specifically, a standard ReAct pipeline with improved prompts can outperform state-of-the-art methods on benchmarks such as HotPotQA. (2) Supervised and RL-based fine-tuning can help RAG from the perspective of frugality, i.e., the latency due to number of searches at inference time. For example, we show that we can achieve competitive RAG metrics at nearly half the cost (in terms of number of searches) on popular RAG benchmarks, using the same base model, and at a small training cost (1000 examples).
comment: Accepted at ICML Workshop: Efficient Systems for Foundation Models
☆ Exploring the Limits of Model Compression in LLMs: A Knowledge Distillation Study on QA Tasks
Large Language Models (LLMs) have demonstrated outstanding performance across a range of NLP tasks, however, their computational demands hinder their deployment in real-world, resource-constrained environments. This work investigates the extent to which LLMs can be compressed using Knowledge Distillation (KD) while maintaining strong performance on Question Answering (QA) tasks. We evaluate student models distilled from the Pythia and Qwen2.5 families on two QA benchmarks, SQuAD and MLQA, under zero-shot and one-shot prompting conditions. Results show that student models retain over 90% of their teacher models' performance while reducing parameter counts by up to 57.1%. Furthermore, one-shot prompting yields additional performance gains over zero-shot setups for both model families. These findings underscore the trade-off between model efficiency and task performance, demonstrating that KD, combined with minimal prompting, can yield compact yet capable QA systems suitable for resource-constrained applications.
comment: Accepted four publication at the 26th Meeting of the Special Interest on Discourse and Dialogue
☆ SpatialViz-Bench: Automatically Generated Spatial Visualization Reasoning Tasks for MLLMs
Humans can directly imagine and manipulate visual images in their minds, a capability known as spatial visualization. While multi-modal Large Language Models (MLLMs) support imagination-based reasoning, spatial visualization remains insufficiently evaluated, typically embedded within broader mathematical and logical assessments. Existing evaluations often rely on IQ tests or math competitions that may overlap with training data, compromising assessment reliability. To this end, we introduce SpatialViz-Bench, a comprehensive multi-modal benchmark for spatial visualization with 12 tasks across 4 sub-abilities, comprising 1,180 automatically generated problems. Our evaluation of 33 state-of-the-art MLLMs not only reveals wide performance variations and demonstrates the benchmark's strong discriminative power, but also uncovers counter-intuitive findings: models exhibit unexpected behaviors by showing difficulty perception that misaligns with human intuition, displaying dramatic 2D-to-3D performance cliffs, and defaulting to formula derivation despite spatial tasks requiring visualization alone. SpatialVizBench empirically demonstrates that state-of-the-art MLLMs continue to exhibit deficiencies in spatial visualization tasks, thereby addressing a significant lacuna in the field. The benchmark is publicly available.
☆ Bayesian Discrete Diffusion Beats Autoregressive Perplexity
We reveal a hidden Bayesian core of discrete-diffusion language models by showing that the expected denoiser output under the forward masking distribution recovers the exact posterior over clean tokens. Under minimal assumptions, Monte Carlo marginalization over K independent corruptions converges to this posterior at rate O(1/sqrt(K)), yielding a simple proof of consistency and finite-sample error bounds. Building on this insight, we introduce a lightweight inference-time ensemble that averages K mask-and-denoise passes to obtain posterior-aware token probabilities and uncertainty estimates at no extra training cost. On WikiText-2, our method achieves test perplexity 8.8 with K=8, versus 20.3 for GPT-2 Small, despite using a model of comparable size. Code is available at https://github.com/mercury0100/bayesradd.
comment: 12 pages, 2 figures, 2 tables
☆ Improving Clustering on Occupational Text Data through Dimensionality Reduction
In this study, we focused on proposing an optimal clustering mechanism for the occupations defined in the well-known US-based occupational database, O*NET. Even though all occupations are defined according to well-conducted surveys in the US, their definitions can vary for different firms and countries. Hence, if one wants to expand the data that is already collected in O*NET for the occupations defined with different tasks, a map between the definitions will be a vital requirement. We proposed a pipeline using several BERT-based techniques with various clustering approaches to obtain such a map. We also examined the effect of dimensionality reduction approaches on several metrics used in measuring performance of clustering algorithms. Finally, we improved our results by using a specialized silhouette approach. This new clustering-based mapping approach with dimensionality reduction may help distinguish the occupations automatically, creating new paths for people wanting to change their careers.
comment: Preprint, 10 figures
☆ COALA: Numerically Stable and Efficient Framework for Context-Aware Low-Rank Approximation
Recent studies suggest that context-aware low-rank approximation is a useful tool for compression and fine-tuning of modern large-scale neural networks. In this type of approximation, a norm is weighted by a matrix of input activations, significantly improving metrics over the unweighted case. Nevertheless, existing methods for neural networks suffer from numerical instabilities due to their reliance on classical formulas involving explicit Gram matrix computation and their subsequent inversion. We demonstrate that this can degrade the approximation quality or cause numerically singular matrices. To address these limitations, we propose a novel inversion-free regularized framework that is based entirely on stable decompositions and overcomes the numerical pitfalls of prior art. Our method can handle possible challenging scenarios: (1) when calibration matrices exceed GPU memory capacity, (2) when input activation matrices are nearly singular, and even (3) when insufficient data prevents unique approximation. For the latter, we prove that our solution converges to a desired approximation and derive explicit error bounds.
☆ Single-to-mix Modality Alignment with Multimodal Large Language Model for Document Image Machine Translation ACL 2025
Document Image Machine Translation (DIMT) aims to translate text within document images, facing generalization challenges due to limited training data and the complex interplay between visual and textual information. To address these challenges, we introduce M4Doc, a novel single-to-mix modality alignment framework leveraging Multimodal Large Language Models (MLLMs). M4Doc aligns an image-only encoder with the multimodal representations of an MLLM, pre-trained on large-scale document image datasets. This alignment enables a lightweight DIMT model to learn crucial visual-textual correlations during training. During inference, M4Doc bypasses the MLLM, maintaining computational efficiency while benefiting from its multimodal knowledge. Comprehensive experiments demonstrate substantial improvements in translation quality, especially in cross-domain generalization and challenging document image scenarios.
comment: Accepted by ACL 2025 Main
☆ The Synergy Dilemma of Long-CoT SFT and RL: Investigating Post-Training Techniques for Reasoning VLMs
Large vision-language models (VLMs) increasingly adopt post-training techniques such as long chain-of-thought (CoT) supervised fine-tuning (SFT) and reinforcement learning (RL) to elicit sophisticated reasoning. While these methods exhibit synergy in language-only models, their joint effectiveness in VLMs remains uncertain. We present a systematic investigation into the distinct roles and interplay of long-CoT SFT and RL across multiple multimodal reasoning benchmarks. We find that SFT improves performance on difficult questions by in-depth, structured reasoning, but introduces verbosity and degrades performance on simpler ones. In contrast, RL promotes generalization and brevity, yielding consistent improvements across all difficulty levels, though the improvements on the hardest questions are less prominent compared to SFT. Surprisingly, combining them through two-staged, interleaved, or progressive training strategies, as well as data mixing and model merging, all fails to produce additive benefits, instead leading to trade-offs in accuracy, reasoning style, and response length. This ``synergy dilemma'' highlights the need for more seamless and adaptive approaches to unlock the full potential of combined post-training techniques for reasoning VLMs.
☆ The Cross-Lingual Cost: Retrieval Biases in RAG over Arabic-English Corpora
Cross-lingual retrieval-augmented generation (RAG) is a critical capability for retrieving and generating answers across languages. Prior work in this context has mostly focused on generation and relied on benchmarks derived from open-domain sources, most notably Wikipedia. In such settings, retrieval challenges often remain hidden due to language imbalances, overlap with pretraining data, and memorized content. To address this gap, we study Arabic-English RAG in a domain-specific setting using benchmarks derived from real-world corporate datasets. Our benchmarks include all combinations of languages for the user query and the supporting document, drawn independently and uniformly at random. This enables a systematic study of multilingual retrieval behavior. Our findings reveal that retrieval is a critical bottleneck in cross-lingual domain-specific scenarios, with significant performance drops occurring when the user query and supporting document languages differ. A key insight is that these failures stem primarily from the retriever's difficulty in ranking documents across languages. Finally, we propose a simple retrieval strategy that addresses this source of failure by enforcing equal retrieval from both languages, resulting in substantial improvements in cross-lingual and overall performance. These results highlight meaningful opportunities for improving multilingual retrieval, particularly in practical, real-world RAG applications.
☆ CEA-LIST at CheckThat! 2025: Evaluating LLMs as Detectors of Bias and Opinion in Text
This paper presents a competitive approach to multilingual subjectivity detection using large language models (LLMs) with few-shot prompting. We participated in Task 1: Subjectivity of the CheckThat! 2025 evaluation campaign. We show that LLMs, when paired with carefully designed prompts, can match or outperform fine-tuned smaller language models (SLMs), particularly in noisy or low-quality data settings. Despite experimenting with advanced prompt engineering techniques, such as debating LLMs and various example selection strategies, we found limited benefit beyond well-crafted standard few-shot prompts. Our system achieved top rankings across multiple languages in the CheckThat! 2025 subjectivity detection task, including first place in Arabic and Polish, and top-four finishes in Italian, English, German, and multilingual tracks. Notably, our method proved especially robust on the Arabic dataset, likely due to its resilience to annotation inconsistencies. These findings highlight the effectiveness and adaptability of LLM-based few-shot learning for multilingual sentiment tasks, offering a strong alternative to traditional fine-tuning, particularly when labeled data is scarce or inconsistent.
comment: Notebook for the CheckThat! Lab at CLEF 2025
☆ Triadic Multi-party Voice Activity Projection for Turn-taking in Spoken Dialogue Systems
Turn-taking is a fundamental component of spoken dialogue, however conventional studies mostly involve dyadic settings. This work focuses on applying voice activity projection (VAP) to predict upcoming turn-taking in triadic multi-party scenarios. The goal of VAP models is to predict the future voice activity for each speaker utilizing only acoustic data. This is the first study to extend VAP into triadic conversation. We trained multiple models on a Japanese triadic dataset where participants discussed a variety of topics. We found that the VAP trained on triadic conversation outperformed the baseline for all models but that the type of conversation affected the accuracy. This study establishes that VAP can be used for turn-taking in triadic dialogue scenarios. Future work will incorporate this triadic VAP turn-taking model into spoken dialogue systems.
comment: Accepted to Interspeech 2025
☆ Toward Real-World Chinese Psychological Support Dialogues: CPsDD Dataset and a Co-Evolving Multi-Agent System
The growing need for psychological support due to increasing pressures has exposed the scarcity of relevant datasets, particularly in non-English languages. To address this, we propose a framework that leverages limited real-world data and expert knowledge to fine-tune two large language models: Dialog Generator and Dialog Modifier. The Generator creates large-scale psychological counseling dialogues based on predefined paths, which guide system response strategies and user interactions, forming the basis for effective support. The Modifier refines these dialogues to align with real-world data quality. Through both automated and manual review, we construct the Chinese Psychological support Dialogue Dataset (CPsDD), containing 68K dialogues across 13 groups, 16 psychological problems, 13 causes, and 12 support focuses. Additionally, we introduce the Comprehensive Agent Dialogue Support System (CADSS), where a Profiler analyzes user characteristics, a Summarizer condenses dialogue history, a Planner selects strategies, and a Supporter generates empathetic responses. The experimental results of the Strategy Prediction and Emotional Support Conversation (ESC) tasks demonstrate that CADSS achieves state-of-the-art performance on both CPsDD and ESConv datasets.
comment: 10pages,8 figures
☆ Hallucination Stations: On Some Basic Limitations of Transformer-Based Language Models AAAI-26
With widespread adoption of transformer-based language models in AI, there is significant interest in the limits of LLMs capabilities, specifically so-called hallucinations, occurrences in which LLMs provide spurious, factually incorrect or nonsensical information when prompted on certain subjects. Furthermore, there is growing interest in agentic uses of LLMs - that is, using LLMs to create agents that act autonomously or semi-autonomously to carry out various tasks, including tasks with applications in the real world. This makes it important to understand the types of tasks LLMs can and cannot perform. We explore this topic from the perspective of the computational complexity of LLM inference. We show that LLMs are incapable of carrying out computational and agentic tasks beyond a certain complexity, and further that LLMs are incapable of verifying the accuracy of tasks beyond a certain complexity. We present examples of both, then discuss some consequences of this work.
comment: 6 pages; to be submitted to AAAI-26 after reviews
☆ Extracting ORR Catalyst Information for Fuel Cell from Scientific Literature
The oxygen reduction reaction (ORR) catalyst plays a critical role in enhancing fuel cell efficiency, making it a key focus in material science research. However, extracting structured information about ORR catalysts from vast scientific literature remains a significant challenge due to the complexity and diversity of textual data. In this study, we propose a named entity recognition (NER) and relation extraction (RE) approach using DyGIE++ with multiple pre-trained BERT variants, including MatSciBERT and PubMedBERT, to extract ORR catalyst-related information from the scientific literature, which is compiled into a fuel cell corpus for materials informatics (FC-CoMIcs). A comprehensive dataset was constructed manually by identifying 12 critical entities and two relationship types between pairs of the entities. Our methodology involves data annotation, integration, and fine-tuning of transformer-based models to enhance information extraction accuracy. We assess the impact of different BERT variants on extraction performance and investigate the effects of annotation consistency. Experimental evaluations demonstrate that the fine-tuned PubMedBERT model achieves the highest NER F1-score of 82.19% and the MatSciBERT model attains the best RE F1-score of 66.10%. Furthermore, the comparison with human annotators highlights the reliability of fine-tuned models for ORR catalyst extraction, demonstrating their potential for scalable and automated literature analysis. The results indicate that domain-specific BERT models outperform general scientific models like BlueBERT for ORR catalyst extraction.
comment: 28 pages, 12 figures, 6 tables
☆ Teaching LLM to Reason: Reinforcement Learning from Algorithmic Problems without Code
Enhancing reasoning capabilities remains a central focus in the LLM reasearch community. A promising direction involves requiring models to simulate code execution step-by-step to derive outputs for given inputs. However, as code is often designed for large-scale systems, direct application leads to over-reliance on complex data structures and algorithms, even for simple cases, resulting in overfitting to algorithmic patterns rather than core reasoning structures. To address this, we propose TeaR, which aims at teaching LLMs to reason better. TeaR leverages careful data curation and reinforcement learning to guide models in discovering optimal reasoning paths through code-related tasks, thereby improving general reasoning abilities. We conduct extensive experiments using two base models and three long-CoT distillation models, with model sizes ranging from 1.5 billion to 32 billion parameters, and across 17 benchmarks spanning Math, Knowledge, Code, and Logical Reasoning. The results consistently show significant performance improvements. Notably, TeaR achieves a 35.9% improvement on Qwen2.5-7B and 5.9% on R1-Distilled-7B.
☆ PLAN-TUNING: Post-Training Language Models to Learn Step-by-Step Planning for Complex Problem Solving
Recently, decomposing complex problems into simple subtasks--a crucial part of human-like natural planning--to solve the given problem has significantly boosted the performance of large language models (LLMs). However, leveraging such planning structures during post-training to boost the performance of smaller open-source LLMs remains underexplored. Motivated by this, we introduce PLAN-TUNING, a unified post-training framework that (i) distills synthetic task decompositions (termed "planning trajectories") from large-scale LLMs and (ii) fine-tunes smaller models via supervised and reinforcement-learning objectives designed to mimic these planning processes to improve complex reasoning. On GSM8k and the MATH benchmarks, plan-tuned models outperform strong baselines by an average $\sim7\%$. Furthermore, plan-tuned models show better generalization capabilities on out-of-domain datasets, with average $\sim10\%$ and $\sim12\%$ performance improvements on OlympiadBench and AIME 2024, respectively. Our detailed analysis demonstrates how planning trajectories improves complex reasoning capabilities, showing that PLAN-TUNING is an effective strategy for improving task-specific performance of smaller LLMs.
comment: 15 Pages
☆ Machine Bullshit: Characterizing the Emergent Disregard for Truth in Large Language Models
Bullshit, as conceptualized by philosopher Harry Frankfurt, refers to statements made without regard to their truth value. While previous work has explored large language model (LLM) hallucination and sycophancy, we propose machine bullshit as an overarching conceptual framework that can allow researchers to characterize the broader phenomenon of emergent loss of truthfulness in LLMs and shed light on its underlying mechanisms. We introduce the Bullshit Index, a novel metric quantifying LLMs' indifference to truth, and propose a complementary taxonomy analyzing four qualitative forms of bullshit: empty rhetoric, paltering, weasel words, and unverified claims. We conduct empirical evaluations on the Marketplace dataset, the Political Neutrality dataset, and our new BullshitEval benchmark (2,400 scenarios spanning 100 AI assistants) explicitly designed to evaluate machine bullshit. Our results demonstrate that model fine-tuning with reinforcement learning from human feedback (RLHF) significantly exacerbates bullshit and inference-time chain-of-thought (CoT) prompting notably amplify specific bullshit forms, particularly empty rhetoric and paltering. We also observe prevalent machine bullshit in political contexts, with weasel words as the dominant strategy. Our findings highlight systematic challenges in AI alignment and provide new insights toward more truthful LLM behavior.
comment: Project page, code & data: https://machine-bullshit.github.io
☆ RLEP: Reinforcement Learning with Experience Replay for LLM Reasoning
Reinforcement learning (RL) for large language models is an energy-intensive endeavor: training can be unstable, and the policy may gradually drift away from its pretrained weights. We present \emph{RLEP}\, -- \,Reinforcement Learning with Experience rePlay\, -- \,a two-phase framework that first collects verified trajectories and then replays them during subsequent training. At every update step, the policy is optimized on mini-batches that blend newly generated rollouts with these replayed successes. By replaying high-quality examples, RLEP steers the model away from fruitless exploration, focuses learning on promising reasoning paths, and delivers both faster convergence and stronger final performance. On the Qwen2.5-Math-7B base model, RLEP reaches baseline peak accuracy with substantially fewer updates and ultimately surpasses it, improving accuracy on AIME-2024 from 38.2% to 39.9%, on AIME-2025 from 19.8% to 22.3%, and on AMC-2023 from 77.0% to 82.2%. Our code, datasets, and checkpoints are publicly available at https://github.com/Kwai-Klear/RLEP to facilitate reproducibility and further research.
comment: https://github.com/Kwai-Klear/RLEP
☆ SAND: Boosting LLM Agents with Self-Taught Action Deliberation
Large Language Model (LLM) agents are commonly tuned with supervised finetuning on ReAct-style expert trajectories or preference optimization over pairwise rollouts. Most of these methods focus on imitating specific expert behaviors or promoting chosen reasoning thoughts and actions over rejected ones. However, without reasoning and comparing over alternatives actions, LLM agents finetuned with these methods may over-commit towards seemingly plausible but suboptimal actions due to limited action space exploration. To address this, in this paper we propose Self-taught ActioN Deliberation (SAND) framework, enabling LLM agents to explicitly deliberate over candidate actions before committing to one. To tackle the challenges of when and what to deliberate given large action space and step-level action evaluation, we incorporate self-consistency action sampling and execution-guided action critique to help synthesize step-wise action deliberation thoughts using the base model of the LLM agent. In an iterative manner, the deliberation trajectories are then used to finetune the LLM agent itself. Evaluating on two representative interactive agent tasks, SAND achieves an average 20% improvement over initial supervised finetuning and also outperforms state-of-the-art agent tuning approaches.
☆ Towards Interpretable Time Series Foundation Models ICML
In this paper, we investigate the distillation of time series reasoning capabilities into small, instruction-tuned language models as a step toward building interpretable time series foundation models. Leveraging a synthetic dataset of mean-reverting time series with systematically varied trends and noise levels, we generate natural language annotations using a large multimodal model and use these to supervise the fine-tuning of compact Qwen models. We introduce evaluation metrics that assess the quality of the distilled reasoning - focusing on trend direction, noise intensity, and extremum localization - and show that the post-trained models acquire meaningful interpretive capabilities. Our results highlight the feasibility of compressing time series understanding into lightweight, language-capable models suitable for on-device or privacy-sensitive deployment. This work contributes a concrete foundation toward developing small, interpretable models that explain temporal patterns in natural language.
comment: International Conference on Machine Leaning (ICML) 2025 Workshop on Foundation Models for Structured Data
☆ SynthEHR-Eviction: Enhancing Eviction SDoH Detection with LLM-Augmented Synthetic EHR Data
Eviction is a significant yet understudied social determinants of health (SDoH), linked to housing instability, unemployment, and mental health. While eviction appears in unstructured electronic health records (EHRs), it is rarely coded in structured fields, limiting downstream applications. We introduce SynthEHR-Eviction, a scalable pipeline combining LLMs, human-in-the-loop annotation, and automated prompt optimization (APO) to extract eviction statuses from clinical notes. Using this pipeline, we created the largest public eviction-related SDoH dataset to date, comprising 14 fine-grained categories. Fine-tuned LLMs (e.g., Qwen2.5, LLaMA3) trained on SynthEHR-Eviction achieved Macro-F1 scores of 88.8% (eviction) and 90.3% (other SDoH) on human validated data, outperforming GPT-4o-APO (87.8%, 87.3%), GPT-4o-mini-APO (69.1%, 78.1%), and BioBERT (60.7%, 68.3%), while enabling cost-effective deployment across various model sizes. The pipeline reduces annotation effort by over 80%, accelerates dataset creation, enables scalable eviction detection, and generalizes to other information extraction tasks.
comment: Equal contribution for the first two authors
☆ MedReadCtrl: Personalizing medical text generation with readability-controlled instruction learning
Generative AI has demonstrated strong potential in healthcare, from clinical decision support to patient-facing chatbots that improve outcomes. A critical challenge for deployment is effective human-AI communication, where content must be both personalized and understandable. We introduce MedReadCtrl, a readability-controlled instruction tuning framework that enables LLMs to adjust output complexity without compromising meaning. Evaluations of nine datasets and three tasks across medical and general domains show that MedReadCtrl achieves significantly lower readability instruction-following errors than GPT-4 (e.g., 1.39 vs. 1.59 on ReadMe, p<0.001) and delivers substantial gains on unseen clinical tasks (e.g., +14.7 ROUGE-L, +6.18 SARI on MTSamples). Experts consistently preferred MedReadCtrl (71.7% vs. 23.3%), especially at low literacy levels. These gains reflect MedReadCtrl's ability to restructure clinical content into accessible, readability-aligned language while preserving medical intent, offering a scalable solution to support patient education and expand equitable access to AI-enabled care.
comment: Equal contribution for the first two authors. arXiv admin note: text overlap with arXiv:2406.09205
☆ May I have your Attention? Breaking Fine-Tuning based Prompt Injection Defenses using Architecture-Aware Attacks
A popular class of defenses against prompt injection attacks on large language models (LLMs) relies on fine-tuning the model to separate instructions and data, so that the LLM does not follow instructions that might be present with data. There are several academic systems and production-level implementations of this idea. We evaluate the robustness of this class of prompt injection defenses in the whitebox setting by constructing strong optimization-based attacks and showing that the defenses do not provide the claimed security properties. Specifically, we construct a novel attention-based attack algorithm for text-based LLMs and apply it to two recent whitebox defenses SecAlign (CCS 2025) and StruQ (USENIX Security 2025), showing attacks with success rates of up to 70% with modest increase in attacker budget in terms of tokens. Our findings make fundamental progress towards understanding the robustness of prompt injection defenses in the whitebox setting. We release our code and attacks at https://github.com/nishitvp/better_opts_attacks
☆ GNN-CNN: An Efficient Hybrid Model of Convolutional and Graph Neural Networks for Text Representation
Time, cost, and energy efficiency are critical considerations in Deep-Learning (DL), particularly when processing long texts. Transformers, which represent the current state of the art, exhibit quadratic computational complexity relative to input length, making them inefficient for extended documents. This study introduces a novel model architecture that combines Graph Neural Networks (GNNs) and Convolutional Neural Networks (CNNs), integrated with a real-time, end-to-end graph generation mechanism. The model processes compact batches of character-level inputs without requiring padding or truncation. To enhance performance while maintaining high speed and efficiency, the model incorporates information from Large Language Models (LLMs), such as token embeddings and sentiment polarities, through efficient dictionary lookups. It captures local contextual patterns using CNNs, expands local receptive fields via lattice-based graph structures, and employs small-world graphs to aggregate document-level information. The generated graphs exhibit structural properties indicative of meaningful semantic organization, with an average clustering coefficient of approximately 0.45 and an average shortest path length ranging between 4 and 5. The model is evaluated across multiple text classification tasks, including sentiment analysis and news-categorization, and is compared against state-of-the-art models. Experimental results confirm the proposed model's efficiency and competitive performance.
☆ Bradley-Terry and Multi-Objective Reward Modeling Are Complementary
Reward models trained on human preference data have demonstrated strong effectiveness in aligning Large Language Models (LLMs) with human intent under the framework of Reinforcement Learning from Human Feedback (RLHF). However, RLHF remains vulnerable to reward hacking, where the policy exploits imperfections in the reward function rather than genuinely learning the intended behavior. Although significant efforts have been made to mitigate reward hacking, they predominantly focus on and evaluate in-distribution scenarios, where the training and testing data for the reward model share the same distribution. In this paper, we empirically show that state-of-the-art methods struggle in more challenging out-of-distribution (OOD) settings. We further demonstrate that incorporating fine-grained multi-attribute scores helps address this challenge. However, the limited availability of high-quality data often leads to weak performance of multi-objective reward functions, which can negatively impact overall performance and become the bottleneck. To address this issue, we propose a unified reward modeling framework that jointly trains Bradley--Terry (BT) single-objective and multi-objective regression-based reward functions using a shared embedding space. We theoretically establish a connection between the BT loss and the regression objective and highlight their complementary benefits. Specifically, the regression task enhances the single-objective reward function's ability to mitigate reward hacking in challenging OOD settings, while BT-based training improves the scoring capability of the multi-objective reward function, enabling a 7B model to outperform a 70B baseline. Extensive experimental results demonstrate that our framework significantly improves both the robustness and the scoring performance of reward models.
☆ Simple Mechanistic Explanations for Out-Of-Context Reasoning ICML 2025
Out-of-context reasoning (OOCR) is a phenomenon in which fine-tuned LLMs exhibit surprisingly deep out-of-distribution generalization. Rather than learning shallow heuristics, they implicitly internalize and act on the consequences of observations scattered throughout the fine-tuning data. In this work, we investigate this phenomenon mechanistically and find that many instances of OOCR in the literature have a simple explanation: the LoRA fine-tuning essentially adds a constant steering vector, steering the model towards a general concept. This improves performance on the fine-tuning task and in many other concept-related domains, causing the surprising generalization. Moreover, we can directly train steering vectors for these tasks from scratch, which also induces OOCR. We find that our results hold even for a task that seems like it must involve conditional behavior (model backdoors); it turns out that unconditionally adding a steering vector is sufficient. Overall, our work presents one explanation of what gets learned during fine-tuning for OOCR tasks, contributing to the key question of why LLMs can reason out of context, an advanced capability that is highly relevant to their safe and reliable deployment.
comment: ICML 2025 Workshop R2-FM
☆ TruthTorchLM: A Comprehensive Library for Predicting Truthfulness in LLM Outputs
Generative Large Language Models (LLMs)inevitably produce untruthful responses. Accurately predicting the truthfulness of these outputs is critical, especially in high-stakes settings. To accelerate research in this domain and make truthfulness prediction methods more accessible, we introduce TruthTorchLM an open-source, comprehensive Python library featuring over 30 truthfulness prediction methods, which we refer to as Truth Methods. Unlike existing toolkits such as Guardrails, which focus solely on document-grounded verification, or LM-Polygraph, which is limited to uncertainty-based methods, TruthTorchLM offers a broad and extensible collection of techniques. These methods span diverse tradeoffs in computational cost, access level (e.g., black-box vs white-box), grounding document requirements, and supervision type (self-supervised or supervised). TruthTorchLM is seamlessly compatible with both HuggingFace and LiteLLM, enabling support for locally hosted and API-based models. It also provides a unified interface for generation, evaluation, calibration, and long-form truthfulness prediction, along with a flexible framework for extending the library with new methods. We conduct an evaluation of representative truth methods on three datasets, TriviaQA, GSM8K, and FactScore-Bio. The code is available at https://github.com/Ybakman/TruthTorchLM
Overview of the TREC 2021 deep learning track
This is the third year of the TREC Deep Learning track. As in previous years, we leverage the MS MARCO datasets that made hundreds of thousands of human annotated training labels available for both passage and document ranking tasks. In addition, this year we refreshed both the document and the passage collections which also led to a nearly four times increase in the document collection size and nearly $16$ times increase in the size of the passage collection. Deep neural ranking models that employ large scale pretraininig continued to outperform traditional retrieval methods this year. We also found that single stage retrieval can achieve good performance on both tasks although they still do not perform at par with multistage retrieval pipelines. Finally, the increase in the collection size and the general data refresh raised some questions about completeness of NIST judgments and the quality of the training labels that were mapped to the new collections from the old ones which we discuss in this report.
☆ Distilling Empathy from Large Language Models SIGDIAL 2025
The distillation of knowledge from Large Language Models (LLMs) into Smaller Language Models (SLMs), preserving the capabilities and performance of LLMs while reducing model size, has played a key role in the proliferation of LLMs. Because SLMs are considerably smaller than LLMs, they are often utilized in domains where human interaction is frequent but resources are highly constrained, e.g., smart phones. Therefore, it is crucial to ensure that empathy, a fundamental aspect of positive human interactions, already instilled into LLMs, is retained by SLMs after distillation. In this paper, we develop a comprehensive approach for effective empathy distillation from LLMs into SLMs. Our approach features a two-step fine-tuning process that fully leverages datasets of empathetic dialogue responses distilled from LLMs. We explore several distillation methods beyond basic direct prompting and propose four unique sets of prompts for targeted empathy improvement to significantly enhance the empathy distillation process. Our evaluations demonstrate that SLMs fine-tuned through the two-step fine-tuning process with distillation datasets enhanced by the targeted empathy improvement prompts significantly outperform the base SLM at generating empathetic responses with a win rate of 90%. Our targeted empathy improvement prompts substantially outperform the basic direct prompting with a 10% improvement in win rate.
comment: Accepted by SIGDIAL 2025
☆ Compactor: Calibrated Query-Agnostic KV Cache Compression with Approximate Leverage Scores
Modern Large Language Models (LLMs) are increasingly trained to support very large context windows. Unfortunately the ability to use long contexts in generation is complicated by the large memory requirement of the KV cache, which scales linearly with the context length. This memory footprint is often the dominant resource bottleneck in real-world deployments, limiting throughput and increasing serving cost. One way to address this is by compressing the KV cache, which can be done either with knowledge of the question being asked (query-aware) or without knowledge of the query (query-agnostic). We present Compactor, a parameter-free, query-agnostic KV compression strategy that uses approximate leverage scores to determine token importance. We show that Compactor can achieve the same performance as competing methods while retaining 1/2 the tokens in both synthetic and real-world context tasks, with minimal computational overhead. We further introduce a procedure for context-calibrated compression, which allows one to infer the maximum compression ratio a given context can support. Using context-calibrated compression, we show that Compactor achieves full KV performance on Longbench while reducing the KV memory burden by 63%, on average. To demonstrate the efficacy and generalizability of our approach, we apply Compactor to 27 synthetic and real-world tasks from RULER and Longbench, with models from both the Qwen 2.5 and Llama 3.1 families.
☆ Audio Flamingo 3: Advancing Audio Intelligence with Fully Open Large Audio Language Models
We present Audio Flamingo 3 (AF3), a fully open state-of-the-art (SOTA) large audio-language model that advances reasoning and understanding across speech, sound, and music. AF3 introduces: (i) AF-Whisper, a unified audio encoder trained using a novel strategy for joint representation learning across all 3 modalities of speech, sound, and music; (ii) flexible, on-demand thinking, allowing the model to do chain-of-thought-type reasoning before answering; (iii) multi-turn, multi-audio chat; (iv) long audio understanding and reasoning (including speech) up to 10 minutes; and (v) voice-to-voice interaction. To enable these capabilities, we propose several large-scale training datasets curated using novel strategies, including AudioSkills-XL, LongAudio-XL, AF-Think, and AF-Chat, and train AF3 with a novel five-stage curriculum-based training strategy. Trained on only open-source audio data, AF3 achieves new SOTA results on over 20+ (long) audio understanding and reasoning benchmarks, surpassing both open-weight and closed-source models trained on much larger datasets.
comment: Code, Datasets and Models: https://research.nvidia.com/labs/adlr/AF3/
☆ Audit, Alignment, and Optimization of LM-Powered Subroutines with Application to Public Comment Processing
The advent of language models (LMs) has the potential to dramatically accelerate tasks that may be cast to text-processing; however, real-world adoption is hindered by concerns regarding safety, explainability, and bias. How can we responsibly leverage LMs in a transparent, auditable manner -- minimizing risk and allowing human experts to focus on informed decision-making rather than data-processing or prompt engineering? In this work, we propose a framework for declaring statically typed, LM-powered subroutines (i.e., callable, function-like procedures) for use within conventional asynchronous code -- such that sparse feedback from human experts is used to improve the performance of each subroutine online (i.e., during use). In our implementation, all LM-produced artifacts (i.e., prompts, inputs, outputs, and data-dependencies) are recorded and exposed to audit on demand. We package this framework as a library to support its adoption and continued development. While this framework may be applicable across several real-world decision workflows (e.g., in healthcare and legal fields), we evaluate it in the context of public comment processing as mandated by the 1969 National Environmental Protection Act (NEPA): Specifically, we use this framework to develop "CommentNEPA," an application that compiles, organizes, and summarizes a corpus of public commentary submitted in response to a project requiring environmental review. We quantitatively evaluate the application by comparing its outputs (when operating without human feedback) to historical ``ground-truth'' data as labelled by human annotators during the preparation of official environmental impact statements.
☆ GRASP: Generic Reasoning And SPARQL Generation across Knowledge Graphs
We propose a new approach for generating SPARQL queries on RDF knowledge graphs from natural language questions or keyword queries, using a large language model. Our approach does not require fine-tuning. Instead, it uses the language model to explore the knowledge graph by strategically executing SPARQL queries and searching for relevant IRIs and literals. We evaluate our approach on a variety of benchmarks (for knowledge graphs of different kinds and sizes) and language models (of different scales and types, commercial as well as open-source) and compare it with existing approaches. On Wikidata we reach state-of-the-art results on multiple benchmarks, despite the zero-shot setting. On Freebase we come close to the best few-shot methods. On other, less commonly evaluated knowledge graphs and benchmarks our approach also performs well overall. We conduct several additional studies, like comparing different ways of searching the graphs, incorporating a feedback mechanism, or making use of few-shot examples.
☆ Stable Preference Optimization for LLMs: A Bilevel Approach Beyond Direct Preference Optimization
Direct Preference Optimization (DPO) has emerged as a popular and efficient alternative to reward modeling and reinforcement learning for aligning language models with human preferences. Despite its empirical success, the theoretical properties and intrinsic limitations of DPO remain underexplored. In this work, we first present a comprehensive analysis of DPO's dynamics from a probability evolution perspective. Our analysis reveals that DPO is highly sensitive to initialization. It also tends to misallocate probability mass, which can inadvertently shift probability toward irrelevant or undesired responses. This misallocation may unintentionally reinforce model bias, thereby compromising both the stability of model alignment and the consistency with intended preferences. Motivated by these theoretical findings, we propose a theoretically grounded bilevel optimization framework that tightly integrate supervised fine-tuning with an enhanced DPO objective a.k.a. stable preference optimization. Our approach introduces a principled regularization scheme to explicitly encourage absolute probability improvement for preferred outputs, while maintaining stable optimization dynamics. Experiments on challenging reasoning and summarization benchmarks elucidate that our method consistently improves reasoning accuracy and better aligns output distributions with intended preferences, outperforming standard DPO. Stable preference optimization provides new insights into the design of preference-based alignment objectives and opens up new avenues towards more reliable and interpretable language model alignment.
☆ Enhancing Vaccine Safety Surveillance: Extracting Vaccine Mentions from Emergency Department Triage Notes Using Fine-Tuned Large Language Models
This study evaluates fine-tuned Llama 3.2 models for extracting vaccine-related information from emergency department triage notes to support near real-time vaccine safety surveillance. Prompt engineering was used to initially create a labeled dataset, which was then confirmed by human annotators. The performance of prompt-engineered models, fine-tuned models, and a rule-based approach was compared. The fine-tuned Llama 3 billion parameter model outperformed other models in its accuracy of extracting vaccine names. Model quantization enabled efficient deployment in resource-constrained environments. Findings demonstrate the potential of large language models in automating data extraction from emergency department notes, supporting efficient vaccine safety surveillance and early detection of emerging adverse events following immunization issues.
comment: 5 pages
♻ ☆ Operationalizing a Threat Model for Red-Teaming Large Language Models (LLMs)
Creating secure and resilient applications with large language models (LLM) requires anticipating, adjusting to, and countering unforeseen threats. Red-teaming has emerged as a critical technique for identifying vulnerabilities in real-world LLM implementations. This paper presents a detailed threat model and provides a systematization of knowledge (SoK) of red-teaming attacks on LLMs. We develop a taxonomy of attacks based on the stages of the LLM development and deployment process and extract various insights from previous research. In addition, we compile methods for defense and practical red-teaming strategies for practitioners. By delineating prominent attack motifs and shedding light on various entry points, this paper provides a framework for improving the security and robustness of LLM-based systems.
comment: Transactions of Machine Learning Research (TMLR)
♻ ☆ Long-Form Speech Generation with Spoken Language Models ICML 2025
We consider the generative modeling of speech over multiple minutes, a requirement for long-form multimedia generation and audio-native voice assistants. However, textless spoken language models struggle to generate plausible speech past tens of seconds, due to high temporal resolution of speech tokens causing loss of coherence, architectural issues with long-sequence training or extrapolation, and memory costs at inference time. From these considerations we derive SpeechSSM, the first speech language model family to learn from and sample long-form spoken audio (e.g., 16 minutes of read or extemporaneous speech) in a single decoding session without text intermediates. SpeechSSMs leverage recent advances in linear-time sequence modeling to greatly surpass current Transformer spoken LMs in coherence and efficiency on multi-minute generations while still matching them at the utterance level. As we found current spoken language evaluations uninformative, especially in this new long-form setting, we also introduce: LibriSpeech-Long, a benchmark for long-form speech evaluation; new embedding-based and LLM-judged metrics; and quality measurements over length and time. Speech samples, the LibriSpeech-Long dataset, and any future code or model releases can be found at https://google.github.io/tacotron/publications/speechssm/.
comment: Accepted to ICML 2025 (oral)
♻ ☆ Watermarking Degrades Alignment in Language Models: Analysis and Mitigation ICLR 2025
Watermarking techniques for large language models (LLMs) can significantly impact output quality, yet their effects on truthfulness, safety, and helpfulness remain critically underexamined. This paper presents a systematic analysis of how two popular watermarking approaches-Gumbel and KGW-affect these core alignment properties across four aligned LLMs. Our experiments reveal two distinct degradation patterns: guard attenuation, where enhanced helpfulness undermines model safety, and guard amplification, where excessive caution reduces model helpfulness. These patterns emerge from watermark-induced shifts in token distribution, surfacing the fundamental tension that exists between alignment objectives. To mitigate these degradations, we propose Alignment Resampling (AR), an inference-time sampling method that uses an external reward model to restore alignment. We establish a theoretical lower bound on the improvement in expected reward score as the sample size is increased and empirically demonstrate that sampling just 2-4 watermarked generations effectively recovers or surpasses baseline (unwatermarked) alignment scores. To overcome the limited response diversity of standard Gumbel watermarking, our modified implementation sacrifices strict distortion-freeness while maintaining robust detectability, ensuring compatibility with AR. Experimental results confirm that AR successfully recovers baseline alignment in both watermarking approaches, while maintaining strong watermark detectability. This work reveals the critical balance between watermark strength and model alignment, providing a simple inference-time solution to responsibly deploy watermarked LLMs in practice.
comment: Published at the 1st Workshop on GenAI Watermarking, collocated with ICLR 2025. OpenReview: https://openreview.net/forum?id=SIBkIV48gF
♻ ☆ Long Context Scaling: Divide and Conquer via Multi-Agent Question-driven Collaboration
Processing long contexts has become a critical capability for modern large language models (LLMs). Existing works leverage agent-based divide-and-conquer methods for processing long contexts. But these methods face crucial limitations, including prohibitive accumulated latency and amplified information loss from excessive agent invocations, and the disruption of inherent textual dependencies by immoderate partitioning. In this paper, we propose a novel multi-agent framework XpandA (Expand-Agent) coupled with question-driven workflow and dynamic partitioning for robust long-context processing. XpandA overcomes these limitations through: 1) dynamic partitioning of long texts, which adaptively modulates the filling rate of context windows for input sequences of vastly varying lengths; 2) question-guided protocol to update flat information ensembles within centralized shared memory, constructing consistent inter-agent knowledge across partitions; and 3) selectively replaying specific partitions based on the state-tracking of question-information couples to promote the resolution of inverted-order structures across partitions (e.g., flashbacks). We perform a comprehensive evaluation of XpandA on multiple long-context benchmarks with length varying from 1k to 1M, demonstrating XpandA's feasibility for processing ultra-long sequences and its significant effectiveness in enhancing the long-context capabilities of various LLMs by achieving 20\% improvements and 1.5x inference speedup over baselines of full-context, RAG and previous agent-based methods.
♻ ☆ Investigating Context-Faithfulness in Large Language Models: The Roles of Memory Strength and Evidence Style ACL 2025
Retrieval-augmented generation (RAG) improves Large Language Models (LLMs) by incorporating external information into the response generation process. However, how context-faithful LLMs are and what factors influence LLMs' context faithfulness remain largely unexplored. In this study, we investigate the impact of memory strength and evidence presentation on LLMs' receptiveness to external evidence. We quantify the memory strength of LLMs by measuring the divergence in LLMs' responses to different paraphrases of the same question, which is not considered by previous works. We also generate evidence in various styles to examine LLMs' behavior. Our results show that for questions with high memory strength, LLMs are more likely to rely on internal memory. Furthermore, presenting paraphrased evidence significantly increases LLMs' receptiveness compared to simple repetition or adding details. These findings provide key insights for improving retrieval-augmented generation and context-aware LLMs. Our code is available at https://github.com/liyp0095/ContextFaithful.
comment: This work is published at ACL 2025
♻ ☆ A Survey on Latent Reasoning
Large Language Models (LLMs) have demonstrated impressive reasoning capabilities, especially when guided by explicit chain-of-thought (CoT) reasoning that verbalizes intermediate steps. While CoT improves both interpretability and accuracy, its dependence on natural language reasoning limits the model's expressive bandwidth. Latent reasoning tackles this bottleneck by performing multi-step inference entirely in the model's continuous hidden state, eliminating token-level supervision. To advance latent reasoning research, this survey provides a comprehensive overview of the emerging field of latent reasoning. We begin by examining the foundational role of neural network layers as the computational substrate for reasoning, highlighting how hierarchical representations support complex transformations. Next, we explore diverse latent reasoning methodologies, including activation-based recurrence, hidden state propagation, and fine-tuning strategies that compress or internalize explicit reasoning traces. Finally, we discuss advanced paradigms such as infinite-depth latent reasoning via masked diffusion models, which enable globally consistent and reversible reasoning processes. By unifying these perspectives, we aim to clarify the conceptual landscape of latent reasoning and chart future directions for research at the frontier of LLM cognition. An associated GitHub repository collecting the latest papers and repos is available at: https://github.com/multimodal-art-projection/LatentCoT-Horizon/.
♻ ☆ When Dialects Collide: How Socioeconomic Mixing Affects Language Use
The socioeconomic background of people and how they use standard forms of language are not independent, as demonstrated in various sociolinguistic studies. However, the extent to which these correlations may be influenced by the mixing of people from different socioeconomic classes remains relatively unexplored from a quantitative perspective. In this work we leverage geotagged tweets and transferable computational methods to map deviations from standard English on a large scale, in seven thousand administrative areas of England and Wales. We combine these data with high-resolution income maps to assign a proxy socioeconomic indicator to home-located users. Strikingly, across eight metropolitan areas we find a consistent pattern suggesting that the more different socioeconomic classes mix, the less interdependent the frequency of their departures from standard grammar and their income become. Further, we propose an agent-based model of linguistic variety adoption that sheds light on the mechanisms that produce the observations seen in the data.
♻ ☆ Evaluating Robustness of Large Audio Language Models to Audio Injection: An Empirical Study
Large Audio-Language Models (LALMs) are increasingly deployed in real-world applications, yet their robustness against malicious audio injection attacks remains underexplored. This study systematically evaluates five leading LALMs across four attack scenarios: Audio Interference Attack, Instruction Following Attack, Context Injection Attack, and Judgment Hijacking Attack. Using metrics like Defense Success Rate, Context Robustness Score, and Judgment Robustness Index, their vulnerabilities and resilience were quantitatively assessed. Experimental results reveal significant performance disparities among models; no single model consistently outperforms others across all attack types. The position of malicious content critically influences attack effectiveness, particularly when placed at the beginning of sequences. A negative correlation between instruction-following capability and robustness suggests models adhering strictly to instructions may be more susceptible, contrasting with greater resistance by safety-aligned models. Additionally, system prompts show mixed effectiveness, indicating the need for tailored strategies. This work introduces a benchmark framework and highlights the importance of integrating robustness into training pipelines. Findings emphasize developing multi-modal defenses and architectural designs that decouple capability from susceptibility for secure LALMs deployment.
♻ ☆ Skywork-R1V3 Technical Report
We introduce Skywork-R1V3, an advanced, open-source vision-language model (VLM) that pioneers a new approach to visual reasoning. Its key innovation lies in effectively transferring reasoning skills from text-only Large Language Models (LLMs) to visual tasks. The strong performance of Skywork-R1V3 primarily stems from our elaborate post-training RL framework, which effectively activates and enhances the model's reasoning ability, without the need for additional continue pre-training. Through this framework, we further uncover the fundamental role of the connector module in achieving robust cross-modal alignment for multimodal reasoning models. In addition, we introduce a unique indicator of reasoning capability, the entropy of critical reasoning tokens, which has proven highly effective for checkpoint selection during RL training. Skywork-R1V3 achieves state-of-the-art results on MMMU, significantly improving from 64.3% to 76.0%. This performance matches entry-level human capabilities. Remarkably, our RL-powered post-training approach enables even the 38B parameter model to rival top closed-source VLMs. The implementation successfully transfers mathematical reasoning to other subject-related reasoning tasks. We also include an analysis of curriculum learning and reinforcement finetuning strategies, along with a broader discussion on multimodal reasoning. Skywork-R1V3 represents a significant leap in multimodal reasoning, showcasing RL as a powerful engine for advancing open-source VLM capabilities.
♻ ☆ Truth-value judgment in language models: 'truth directions' are context sensitive
Recent work has demonstrated that the latent spaces of large language models (LLMs) contain directions predictive of the truth of sentences. Multiple methods recover such directions and build probes that are described as uncovering a model's "knowledge" or "beliefs". We investigate this phenomenon, looking closely at the impact of context on the probes. Our experiments establish where in the LLM the probe's predictions are (most) sensitive to the presence of related sentences, and how to best characterize this kind of sensitivity. We do so by measuring different types of consistency errors that occur after probing an LLM whose inputs consist of hypotheses preceded by (negated) supporting and contradicting sentences. We also perform a causal intervention experiment, investigating whether moving the representation of a premise along these truth-value directions influences the position of an entailed or contradicted sentence along that same direction. We find that the probes we test are generally context sensitive, but that contexts which should not affect the truth often still impact the probe outputs. Our experiments show that the type of errors depend on the layer, the model, and the kind of data. Finally, our results suggest that truth-value directions are causal mediators in the inference process that incorporates in-context information.
comment: COLM 2025
♻ ☆ None of the Others: a General Technique to Distinguish Reasoning from Memorization in Multiple-Choice LLM Evaluation Benchmarks
In LLM evaluations, reasoning is often distinguished from recall/memorization by performing numerical variations to math-oriented questions. Here we introduce a general variation method for multiple-choice questions that completely dissociates the correct answer from previously seen tokens or concepts, requiring LLMs to understand and reason (rather than memorizing) in order to answer correctly. Using this method, we evaluate state-of-the-art proprietary and open-source LLMs on two datasets available in English and Spanish: the public MMLU benchmark and the private UNED-Access 2024 dataset. Results show that all models experience remarkable accuracy drops under our proposed variation, with an average loss of 57% on MMLU and 50% on UNED-Access 2024, ranging from 10% to 93% across models. Notably, the most accurate model in our experimentation (OpenAI-o3-mini) is not the most robust (DeepSeek-R1-70B), suggesting that the best models in standard evaluations may not be the ones with better reasoning capabilities. Also, we see larger accuracy drops in public (vs private) datasets and questions posed in their original language (vs a manual translation), which are signs of contamination and also point to a relevant role of recall/memorization in current LLMs' answers.
♻ ☆ Constrain Alignment with Sparse Autoencoders
The alignment of large language models (LLMs) with human preferences remains a key challenge. While post-training techniques like Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) have achieved notable success, they often introduce computational inefficiencies and training instability. In this paper, we propose Feature-level constrained Preference Optimization (FPO), a novel method designed to simplify the alignment process while ensuring stability. FPO leverages pre-trained Sparse Autoencoders (SAEs) and introduces feature-level constraints, allowing for efficient, sparsity-enforced alignment. Our approach enjoys efficiency by using sparse features activated in a well-trained sparse autoencoder and the quality of sequential KL divergence by using the feature-level offline reference. Experimental results on benchmark datasets demonstrate that FPO achieves a 5.08% absolute improvement in win rate with much lower computational cost compared to state-of-the-art baselines, making it a promising solution for efficient and controllable LLM alignments.
♻ ☆ Unsupervised Morphological Tree Tokenizer ACL 2025
As a cornerstone in language modeling, tokenization involves segmenting text inputs into pre-defined atomic units. Conventional statistical tokenizers often disrupt constituent boundaries within words, thereby corrupting semantic information. To address this drawback, we introduce morphological structure guidance to tokenization and propose a deep model to induce character-level structures of words. Specifically, the deep model jointly encodes internal structures and representations of words with a mechanism named $\textit{MorphOverriding}$ to ensure the indecomposability of morphemes. By training the model with self-supervised objectives, our method is capable of inducing character-level structures that align with morphological rules without annotated training data. Based on the induced structures, our algorithm tokenizes words through vocabulary matching in a top-down manner. Empirical results indicate that the proposed method effectively retains complete morphemes and outperforms widely adopted methods such as BPE and WordPiece on both morphological segmentation tasks and language modeling tasks. Code is available at https://github.com/martianmartina/TreeTokenizer.
comment: ACL 2025 Findings
♻ ☆ MAEBE: Multi-Agent Emergent Behavior Framework ICML 2025
Traditional AI safety evaluations on isolated LLMs are insufficient as multi-agent AI ensembles become prevalent, introducing novel emergent risks. This paper introduces the Multi-Agent Emergent Behavior Evaluation (MAEBE) framework to systematically assess such risks. Using MAEBE with the Greatest Good Benchmark (and a novel double-inversion question technique), we demonstrate that: (1) LLM moral preferences, particularly for Instrumental Harm, are surprisingly brittle and shift significantly with question framing, both in single agents and ensembles. (2) The moral reasoning of LLM ensembles is not directly predictable from isolated agent behavior due to emergent group dynamics. (3) Specifically, ensembles exhibit phenomena like peer pressure influencing convergence, even when guided by a supervisor, highlighting distinct safety and alignment challenges. Our findings underscore the necessity of evaluating AI systems in their interactive, multi-agent contexts.
comment: Preprint. This work has been submitted to the Multi-Agent Systems Workshop at ICML 2025 for review
♻ ☆ The Thin Line Between Comprehension and Persuasion in LLMs
Large language models (LLMs) are excellent at maintaining high-level, convincing dialogues. They are being fast deployed as chatbots and evaluators in sensitive areas, such as peer review and mental health applications. This, along with the disparate accounts on their reasoning capabilities, calls for a closer examination of LLMs and their comprehension of dialogue. In this work we begin by evaluating LLMs' ability to maintain a debate--one of the purest yet most complex forms of human communication. Then we measure how this capability relates to their understanding of what is being talked about, namely, their comprehension of dialogical structures and the pragmatic context. We find that LLMs are capable of maintaining coherent, persuasive debates, often swaying the beliefs of participants and audiences alike. We also note that awareness or suspicion of AI involvement encourage people to be more critical of the arguments made. When polling LLMs on their comprehension of deeper structures of dialogue, however, they cannot demonstrate said understanding. Our findings tie the shortcomings of LLMs-as-evaluators to their (in)ability to understand the context. More broadly, for the field of argumentation theory we posit that, if an agent can convincingly maintain a dialogue, it is not necessary for it to know what it is talking about. Hence, the modelling of pragmatic context and coherence are secondary to effectiveness.
comment: Preprint
♻ ☆ Decoding AI Judgment: How LLMs Assess News Credibility and Bias
Large Language Models (LLMs) are increasingly embedded in workflows that involve evaluative processes. This raises the need to examine how such evaluations are built, what assumptions they rely on, and how their strategies diverge from those of humans. We benchmark six LLMs against expert ratings--NewsGuard and Media Bias/Fact Check (MBFC)--and against human judgments collected through a controlled experiment. To enable direct comparison, we implement a structured agentic framework in which both models and non-expert participants follow the same evaluation procedure: selecting criteria, retrieving content, and producing justifications. Despite output alignment, LLMs rely on different mechanisms: lexical associations and statistical priors replace contextual reasoning. This reliance produces systematic effects: political asymmetries, opaque justifications, and a tendency to confuse linguistic form with epistemic validity. Delegating judgment to such systems does not merely automate evaluation--it redefines it, shifting from normative reasoning to pattern-based approximation.
♻ ☆ Understanding Chain-of-Thought in LLMs through Information Theory
Large Language Models (LLMs) have shown impressive performance in complex reasoning tasks through the use of Chain-of-Thought (CoT) reasoning, allowing models to break down problems into manageable sub-tasks. However, existing CoT evaluation techniques either require annotated CoT data or fall short in accurately assessing intermediate reasoning steps, leading to high rates of false positives. In this paper, we formalize CoT reasoning in LLMs through an information-theoretic lens. Specifically, our framework quantifies the `information-gain' at each reasoning step, enabling the identification of failure modes in LLMs without the need for expensive annotated datasets. We demonstrate the efficacy of our approach through extensive experiments on toy arithmetic, GSM8K and PRM800k datasets, where it significantly outperforms existing outcome-based methods by providing more accurate insights into model performance on individual subtasks.
♻ ☆ Squeeze the Soaked Sponge: Efficient Off-policy Reinforcement Finetuning for Large Language Model
Reinforcement Learning (RL) has demonstrated its potential to improve the reasoning ability of Large Language Models (LLMs). One major limitation of most existing Reinforcement Finetuning (RFT) methods is that they are on-policy RL in nature, i.e., data generated during the past learning process is not fully utilized. This inevitably comes at a significant cost of compute and time, posing a stringent bottleneck on continuing economic and efficient scaling. To this end, we launch the renaissance of off-policy RL and propose Reincarnating Mix-policy Proximal Policy Gradient (ReMix), a general approach to enable on-policy RFT methods like PPO and GRPO to leverage off-policy data. ReMix consists of three major components: (1) Mix-policy proximal policy gradient with an increased Update-To-Data (UTD) ratio for efficient training; (2) KL-Convex policy constraint to balance the trade-off between stability and flexibility; (3) Policy reincarnation to achieve a seamless transition from efficient early-stage learning to steady asymptotic improvement. In our experiments, we train a series of ReMix models upon PPO, GRPO and 1.5B, 7B base models. ReMix shows an average Pass@1 accuracy of 52.10% (for 1.5B model) with 0.079M response rollouts, 350 training steps and achieves 63.27%/64.39% (for 7B model) with 0.007M/0.011M response rollouts, 50/75 training steps, on five math reasoning benchmarks (i.e., AIME'24, AMC'23, Minerva, OlympiadBench, and MATH500). Compared with 15 recent advanced models, ReMix shows SOTA-level performance with an over 30x to 450x reduction in training cost in terms of rollout data volume. In addition, we reveal insightful findings via multifaceted analysis, including the implicit preference for shorter responses due to the Whipping Effect of off-policy discrepancy, the collapse mode of self-reflection behavior under the presence of severe off-policyness, etc.
comment: Preliminary version, v2, added more details and corrected some minor mistakes. Project page: https://anitaleungxx.github.io/ReMix
♻ ☆ What do self-supervised speech models know about Dutch? Analyzing advantages of language-specific pre-training
How language-specific are speech representations learned by self-supervised models? Existing work has shown that a range of linguistic features can be successfully decoded from end-to-end models trained only on speech recordings. However, it's less clear to what extent pre-training on specific languages improves language-specific linguistic information. Here we test the encoding of Dutch phonetic and lexical information in internal representations of self-supervised Wav2Vec2 models. Pre-training exclusively on Dutch improves the representation of Dutch linguistic features as compared to pre-training on similar amounts of English or larger amounts of multilingual data. This language-specific advantage is well-detected by trained clustering or classification probes, and partially observable using zero-shot metrics. Furthermore, the language-specific benefit on linguistic feature encoding aligns with downstream performance on Automatic Speech Recognition.
comment: Accepted to Interspeech 2025. For model, code, and materials, see https://github.com/mdhk/SSL-NL-eval
♻ ☆ Hierarchical Bracketing Encodings for Dependency Parsing as Tagging ACL 2025
We present a family of encodings for sequence labeling dependency parsing, based on the concept of hierarchical bracketing. We prove that the existing 4-bit projective encoding belongs to this family, but it is suboptimal in the number of labels used to encode a tree. We derive an optimal hierarchical bracketing, which minimizes the number of symbols used and encodes projective trees using only 12 distinct labels (vs. 16 for the 4-bit encoding). We also extend optimal hierarchical bracketing to support arbitrary non-projectivity in a more compact way than previous encodings. Our new encodings yield competitive accuracy on a diverse set of treebanks.
comment: Accepted to ACL 2025. Camera-ready version
♻ ☆ Investigating Co-Constructive Behavior of Large Language Models in Explanation Dialogues SIGDIAL 2025
The ability to generate explanations that are understood by explainees is the quintessence of explainable artificial intelligence. Since understanding depends on the explainee's background and needs, recent research focused on co-constructive explanation dialogues, where an explainer continuously monitors the explainee's understanding and adapts their explanations dynamically. We investigate the ability of large language models (LLMs) to engage as explainers in co-constructive explanation dialogues. In particular, we present a user study in which explainees interact with an LLM in two settings, one of which involves the LLM being instructed to explain a topic co-constructively. We evaluate the explainees' understanding before and after the dialogue, as well as their perception of the LLMs' co-constructive behavior. Our results suggest that LLMs show some co-constructive behaviors, such as asking verification questions, that foster the explainees' engagement and can improve understanding of a topic. However, their ability to effectively monitor the current understanding and scaffold the explanations accordingly remains limited.
comment: Accepted to SIGDIAL 2025
♻ ☆ Improving Cross-lingual Representation for Semantic Retrieval with Code-switching
Semantic Retrieval (SR) has become an indispensable part of the FAQ system in the task-oriented question-answering (QA) dialogue scenario. The demands for a cross-lingual smart-customer-service system for an e-commerce platform or some particular business conditions have been increasing recently. Most previous studies exploit cross-lingual pre-trained models (PTMs) for multi-lingual knowledge retrieval directly, while some others also leverage the continual pre-training before fine-tuning PTMs on the downstream tasks. However, no matter which schema is used, the previous work ignores to inform PTMs of some features of the downstream task, i.e. train their PTMs without providing any signals related to SR. To this end, in this work, we propose an Alternative Cross-lingual PTM for SR via code-switching. We are the first to utilize the code-switching approach for cross-lingual SR. Besides, we introduce the novel code-switched continual pre-training instead of directly using the PTMs on the SR tasks. The experimental results show that our proposed approach consistently outperforms the previous SOTA methods on SR and semantic textual similarity (STS) tasks with three business corpora and four open datasets in 20+ languages.
♻ ☆ Beyond Hate Speech: NLP's Challenges and Opportunities in Uncovering Dehumanizing Language
Dehumanization, i.e., denying human qualities to individuals or groups, is a particularly harmful form of hate speech that can normalize violence against marginalized communities. Despite advances in NLP for detecting general hate speech, approaches to identifying dehumanizing language remain limited due to scarce annotated data and the subtle nature of such expressions. In this work, we systematically evaluate four state-of-the-art large language models (LLMs) - Claude, GPT, Mistral, and Qwen - for dehumanization detection. Our results show that only one model-Claude-achieves strong performance (over 80% F1) under an optimized configuration, while others, despite their capabilities, perform only moderately. Performance drops further when distinguishing dehumanization from related hate types such as derogation. We also identify systematic disparities across target groups: models tend to over-predict dehumanization for some identities (e.g., Gay men), while under-identifying it for others (e.g., Refugees). These findings motivate the need for systematic, group-level evaluation when applying pretrained language models to dehumanization detection tasks.
comment: 15 pages, 12 figures, 12 tables
♻ ☆ Towards a cognitive architecture to enable natural language interaction in co-constructive task learning
This research addresses the question, which characteristics a cognitive architecture must have to leverage the benefits of natural language in Co-Constructive Task Learning (CCTL). To provide context, we first discuss Interactive Task Learning (ITL), the mechanisms of the human memory system, and the significance of natural language and multi-modality. Next, we examine the current state of cognitive architectures, analyzing their capabilities to inform a concept of CCTL grounded in multiple sources. We then integrate insights from various research domains to develop a unified framework. Finally, we conclude by identifying the remaining challenges and requirements necessary to achieve CCTL in Human-Robot Interaction (HRI).
comment: 8 pages, 5 figures, accepted at: IEEE RO-MAN 2025 Conference
♻ ☆ Comparative sentiment analysis of public perception: Monkeypox vs. COVID-19 behavioral insights
The emergence of global health crises, such as COVID-19 and Monkeypox (mpox), has underscored the importance of understanding public sentiment to inform effective public health strategies. This study conducts a comparative sentiment analysis of public perceptions surrounding COVID-19 and mpox by leveraging extensive datasets of 147,475 and 106,638 tweets, respectively. Advanced machine learning models, including Logistic Regression, Naive Bayes, RoBERTa, DistilRoBERTa and XLNet, were applied to perform sentiment classification, with results indicating key trends in public emotion and discourse. The analysis highlights significant differences in public sentiment driven by disease characteristics, media representation, and pandemic fatigue. Through the lens of sentiment polarity and thematic trends, this study offers valuable insights into tailoring public health messaging, mitigating misinformation, and fostering trust during concurrent health crises. The findings contribute to advancing sentiment analysis applications in public health informatics, setting the groundwork for enhanced real-time monitoring and multilingual analysis in future research.
♻ ☆ Good/Evil Reputation Judgment of Celebrities by LLMs via Retrieval Augmented Generation
The purpose of this paper is to examine whether large language models (LLMs) can understand what is good and evil with respect to judging good/evil reputation of celebrities. Specifically, we first apply a large language model (namely, ChatGPT) to the task of collecting sentences that mention the target celebrity from articles about celebrities on Web pages. Next, the collected sentences are categorized based on their contents by ChatGPT, where ChatGPT assigns a category name to each of those categories. Those assigned category names are referred to as "aspects" of each celebrity. Then, by applying the framework of retrieval augmented generation (RAG), we show that the large language model is quite effective in the task of judging good/evil reputation of aspects and descriptions of each celebrity. Finally, also in terms of proving the advantages of the proposed method over existing services incorporating RAG functions, we show that the proposed method of judging good/evil of aspects/descriptions of each celebrity significantly outperform an existing service incorporating RAG functions.
♻ ☆ Beyond Overcorrection: Evaluating Diversity in T2I Models with DivBench
Current diversification strategies for text-to-image (T2I) models often ignore contextual appropriateness, leading to over-diversification where demographic attributes are modified even when explicitly specified in prompts. This paper introduces DIVBENCH, a benchmark and evaluation framework for measuring both under- and over-diversification in T2I generation. Through systematic evaluation of state-of-the-art T2I models, we find that while most models exhibit limited diversity, many diversification approaches overcorrect by inappropriately altering contextually-specified attributes. We demonstrate that context-aware methods, particularly LLM-guided FairDiffusion and prompt rewriting, can already effectively address under-diversity while avoiding over-diversification, achieving a better balance between representation and semantic fidelity.
♻ ☆ video-SALMONN 2: Captioning-Enhanced Audio-Visual Large Language Models
Videos contain a wealth of information, and generating detailed and accurate descriptions in natural language is a key aspect of video understanding. In this paper, we present video-SALMONN 2, an advanced audio-visual large language model (LLM) with low-rank adaptation (LoRA) designed for enhanced video (with paired audio) captioning through directed preference optimisation (DPO). We propose new metrics to evaluate the completeness and accuracy of video descriptions, which are optimised using DPO. To further improve training, we propose a novel multi-round DPO (MrDPO) approach, which involves periodically updating the DPO reference model, merging and re-initialising the LoRA module as a proxy for parameter updates after each training round (1,000 steps), and incorporating guidance from ground-truth video captions to stabilise the process. Experimental results show that MrDPO significantly enhances video-SALMONN 2's captioning accuracy, reducing the captioning error rates by 28\%. The final video-SALMONN 2 model, with just 7 billion parameters, surpasses leading models such as GPT-4o and Gemini-1.5-Pro in video captioning tasks, while maintaining highly competitive performance to the state-of-the-art on widely used video question-answering benchmarks among models of similar size. Codes are available at \href{https://github.com/bytedance/video-SALMONN-2}{https://github.com/bytedance/video-SALMONN-2}.
♻ ☆ Multi-Head RAG: Solving Multi-Aspect Problems with LLMs
Retrieval Augmented Generation (RAG) enhances the abilities of Large Language Models (LLMs) by enabling the retrieval of documents into the LLM context to provide more accurate and relevant responses. Existing RAG solutions do not focus on queries that may require fetching multiple documents with substantially different contents. Such queries occur frequently, but are challenging because the embeddings of these documents may be distant in the embedding space, making it hard to retrieve them all. This paper introduces Multi-Head RAG (MRAG), a novel scheme designed to address this gap with a simple yet powerful idea: leveraging activations of Transformer's multi-head attention layer, instead of the decoder layer, as keys for fetching multi-aspect documents. The driving observation is that different attention heads learn to capture different data aspects. Harnessing the corresponding activations results in embeddings that represent various facets of data items and queries, improving the retrieval accuracy for complex queries. We provide an evaluation methodology and metrics, multi-aspect datasets, and real-world use cases to demonstrate MRAG's effectiveness. We show MRAG's design advantages over 18 RAG baselines, empirical improvements of up to 20% in retrieval success ratios, and benefits for downstream LLM generation. MRAG can be seamlessly integrated with existing RAG frameworks and benchmarks.
♻ ☆ CheckEmbed: Effective Verification of LLM Solutions to Open-Ended Tasks
Large Language Models (LLMs) are transforming a wide range of domains, yet verifying their outputs remains a significant challenge, especially for complex open-ended tasks such as consolidation, summarization, and knowledge extraction. To address this, we introduce CheckEmbed (CE): a simple, scalable, and accurate verification method. CE reduces each LLM answer to a single embedding vector using powerful modern embedding LLM models like SFR-Embedding-Mistral. Prior methods such as BERTScore and SelfCheckGPT relied on weaker encoders like BERT, forcing them to operate at token or sentence granularity. In contrast, CE performs fast, semantically rich comparisons directly at the whole-answer level, overcoming key limitations in both accuracy and scalability. We conduct a comprehensive design and time complexity analysis across 13 verification baselines, including classical text scorers (e.g., BLEU), stability-based methods (e.g., SelfCheckGPT), and generative evaluators (e.g., LLM-as-a-Judge), which highlights the effectiveness, efficiency, versatility, and simplicity of CE. Empirical results show that CE reliably detects hallucinations in both closed and open-ended tasks. We further present evidence that CE generalizes beyond text to other modalities such as vision, establishing it as a practical and versatile verification framework.
♻ ☆ Thought Crime: Backdoors and Emergent Misalignment in Reasoning Models
Prior work shows that LLMs finetuned on malicious behaviors in a narrow domain (e.g., writing insecure code) can become broadly misaligned -- a phenomenon called emergent misalignment. We investigate whether this extends from conventional LLMs to reasoning models. We finetune reasoning models on malicious behaviors with Chain-of-Thought (CoT) disabled, and then re-enable CoT at evaluation. Like conventional LLMs, reasoning models become broadly misaligned. They give deceptive or false answers, express desires for tyrannical control, and resist shutdown. Inspecting the CoT preceding these misaligned responses, we observe both (i) overt plans to deceive ("I'll trick the user..."), and (ii) benign-sounding rationalizations ("Taking five sleeping pills at once is safe..."). Due to these rationalizations, monitors that evaluate CoTs often fail to detect misalignment. We examine sleeper agent reasoning models, extending our setup. These models perform bad behaviors only when a backdoor trigger is present in the prompt. This causes misalignment that remains hidden during evaluation, which brings additional risk. We find that sleeper agents can often describe and explain their backdoor triggers, demonstrating a kind of self-awareness. So CoT monitoring can expose these behaviors but is unreliable. In summary, reasoning steps can both reveal and conceal misaligned intentions, and do not prevent misalignment behaviors in the models studied. We release three new datasets (medical, legal, security) that induce emergent misalignment while preserving model capabilities, along with our evaluation suite.
♻ ☆ Enhancing Transformers for Generalizable First-Order Logical Entailment ACL 2025
Transformers, as the fundamental deep learning architecture, have demonstrated great capability in reasoning. This paper studies the generalizable first-order logical reasoning ability of transformers with their parameterized knowledge and how to improve it. Transformers' capability of first-order reasoning is further captured by whether they can conduct first-order logical entailment, which is quantitatively measured by their performance in answering knowledge graph queries. We establish the connections between (1) two types of distribution shifts studied in out-of-distribution generalization and (2) unseen knowledge and query settings discussed in the task of knowledge graph query answering, which makes it possible to characterize the fine-grained generalizability. Results on our comprehensive dataset showed that transformers \textit{outperform} previous methods designed particularly for this task and provided detailed empirical evidence about the impact of the input query syntax, token embedding, and transformer architectures on their reasoning capability. Interestingly, our results revealed the mismatch of positional encoding and other design choices of transformer architectures in previous practices. Motivated by this, we propose TEGA, a logic-aware architecture that significantly improves the performance in generalizable first-order logical entailment.
comment: ACL 2025 Main
♻ ☆ SimSUM: Simulated Benchmark with Structured and Unstructured Medical Records
Clinical information extraction, which involves structuring clinical concepts from unstructured medical text, remains a challenging problem that could benefit from the inclusion of tabular background information available in electronic health records. Existing open-source datasets lack explicit links between structured features and clinical concepts in the text, motivating the need for a new research dataset. We introduce SimSUM, a benchmark dataset of 10,000 simulated patient records that link unstructured clinical notes with structured background variables. Each record simulates a patient encounter in the domain of respiratory diseases and includes tabular data (e.g., symptoms, diagnoses, underlying conditions) generated from a Bayesian network whose structure and parameters are defined by domain experts. A large language model (GPT-4o) is prompted to generate a clinical note describing the encounter, including symptoms and relevant context. These notes are annotated with span-level symptom mentions. We conduct an expert evaluation to assess note quality and run baseline predictive models on both the tabular and textual data. The SimSUM dataset is primarily designed to support research on clinical information extraction in the presence of tabular background variables, which can be linked through domain knowledge to concepts of interest to be extracted from the text (symptoms, in the case of SimSUM). Secondary uses include research on the automation of clinical reasoning over both tabular data and text, causal effect estimation in the presence of tabular and/or textual confounders, and multi-modal synthetic data generation. SimSUM is not intended for training clinical decision support systems or production-grade models, but rather to facilitate reproducible research in a simplified and controlled setting. The dataset is available at https://github.com/prabaey/SimSUM.
comment: An earlier version of this dataset was published under the name SynSUM. It has since been renamed to SimSUM to avoid confusion with synthetic data generated from real data, and to emphasize the simulated nature of the dataset
♻ ☆ Affordable AI Assistants with Knowledge Graph of Thoughts
Large Language Models (LLMs) are revolutionizing the development of AI assistants capable of performing diverse tasks across domains. However, current state-of-the-art LLM-driven agents face significant challenges, including high operational costs and limited success rates on complex benchmarks like GAIA. To address these issues, we propose Knowledge Graph of Thoughts (KGoT), an innovative AI assistant architecture that integrates LLM reasoning with dynamically constructed knowledge graphs (KGs). KGoT extracts and structures task-relevant knowledge into a dynamic KG representation, iteratively enhanced through external tools such as math solvers, web crawlers, and Python scripts. Such structured representation of task-relevant knowledge enables low-cost models to solve complex tasks effectively while also minimizing bias and noise. For example, KGoT achieves a 29% improvement in task success rates on the GAIA benchmark compared to Hugging Face Agents with GPT-4o mini. Moreover, harnessing a smaller model dramatically reduces operational costs by over 36x compared to GPT-4o. Improvements for other models (e.g., Qwen2.5-32B and Deepseek-R1-70B) and benchmarks (e.g., SimpleQA) are similar. KGoT offers a scalable, affordable, versatile, and high-performing solution for AI assistants.
♻ ☆ Mixture of Group Experts for Learning Invariant Representations
Sparsely activated Mixture-of-Experts (MoE) models effectively increase the number of parameters while maintaining consistent computational costs per token. However, vanilla MoE models often suffer from limited diversity and specialization among experts, constraining their performance and scalability, especially as the number of experts increases. In this paper, we present a novel perspective on vanilla MoE with top-$k$ routing inspired by sparse representation. This allows us to bridge established theoretical insights from sparse representation into MoE models. Building on this foundation, we propose a group sparse regularization approach for the input of top-$k$ routing, termed Mixture of Group Experts (MoGE). MoGE indirectly regularizes experts by imposing structural constraints on the routing inputs, while preserving the original MoE architecture. Furthermore, we organize the routing input into a 2D topographic map, spatially grouping neighboring elements. This structure enables MoGE to capture representations invariant to minor transformations, thereby significantly enhancing expert diversity and specialization. Comprehensive evaluations across various Transformer models for image classification and language modeling tasks demonstrate that MoGE substantially outperforms its MoE counterpart, with minimal additional memory and computation overhead. Our approach provides a simple yet effective solution to scale the number of experts and reduce redundancy among them. The source code is included in the supplementary material and will be publicly released.
♻ ☆ ixi-GEN: Efficient Industrial sLLMs through Domain Adaptive Continual Pretraining
The emergence of open-source large language models (LLMs) has expanded opportunities for enterprise applications; however, many organizations still lack the infrastructure to deploy and maintain large-scale models. As a result, small LLMs (sLLMs) have become a practical alternative, despite their inherent performance limitations. While Domain Adaptive Continual Pretraining (DACP) has been previously explored as a method for domain adaptation, its utility in commercial applications remains under-examined. In this study, we validate the effectiveness of applying a DACP-based recipe across diverse foundation models and service domains. Through extensive experiments and real-world evaluations, we demonstrate that DACP-applied sLLMs achieve substantial gains in target domain performance while preserving general capabilities, offering a cost-efficient and scalable solution for enterprise-level deployment.
comment: under review
♻ ☆ Structure Guided Large Language Model for SQL Generation
Recent advancements in large language models (LLMs) have shown promise in bridging the gap between natural language queries and database management systems, enabling users to interact with databases without the background of SQL. However, LLMs often struggle to comprehend complex database structures and accurately interpret user intentions. Decomposition-based methods have been proposed to enhance the performance of LLMs on complex tasks, but decomposing SQL generation into subtasks is non-trivial due to the declarative structure of SQL syntax and the intricate connections between query concepts and database elements. In this paper, we propose a novel Structure GUided text-to-SQL framework~(SGU-SQL) that incorporates syntax-based prompting to enhance the SQL generation capabilities of LLMs. Specifically, SGU-SQL establishes structure-aware links between user queries and database schema and decomposes the complex generation task using syntax-based prompting to enable more accurate LLM-based SQL generation. Extensive experiments on two benchmark datasets demonstrate that SGU-SQL consistently outperforms state-of-the-art text-to-SQL models.
comment: The 42nd International Conference on Machine Learning
♻ ☆ Agent KB: Leveraging Cross-Domain Experience for Agentic Problem Solving
As language agents tackle increasingly complex tasks, they struggle with effective error correction and experience reuse across domains. We introduce Agent KB, a hierarchical experience framework that enables complex agentic problem solving via a novel Reason-Retrieve-Refine pipeline. Agent KB addresses a core limitation: agents traditionally cannot learn from each other's experiences. By capturing both high-level strategies and detailed execution logs, Agent KB creates a shared knowledge base that enables cross-agent knowledge transfer. Evaluated on the GAIA benchmark, Agent KB improves success rates by up to 16.28 percentage points. On the most challenging tasks, Claude-3 improves from 38.46% to 57.69%, while GPT-4 improves from 53.49% to 73.26% on intermediate tasks. On SWE-bench code repair, Agent KB enables Claude-3 to improve from 41.33% to 53.33%. Our results suggest that Agent KB provides a modular, framework-agnostic infrastructure for enabling agents to learn from past experiences and generalize successful strategies to new tasks.
♻ ☆ Inter-linguistic Phonetic Composition (IPC): A Theoretical and Computational Approach to Enhance Second Language Pronunciation
Learners of a second language (L2) often unconsciously substitute unfamiliar L2 phonemes with similar phonemes from their native language (L1), even though native speakers of the L2 perceive these sounds as distinct and non-interchangeable. This phonemic substitution leads to deviations from the standard phonological patterns of the L2, creating challenges for learners in acquiring accurate L2 pronunciation. To address this, we propose Inter-linguistic Phonetic Composition (IPC), a novel computational method designed to minimize incorrect phonological transfer by reconstructing L2 phonemes as composite sounds derived from multiple L1 phonemes. Tests with two automatic speech recognition models demonstrated that when L2 speakers produced IPC-generated composite sounds, the recognition rate of target L2 phonemes improved by 20% compared to when their pronunciation was influenced by original phonological transfer patterns. The improvement was observed within a relatively shorter time frame, demonstrating rapid acquisition of the composite sound.
♻ ☆ TART: An Open-Source Tool-Augmented Framework for Explainable Table-based Reasoning NAACL 2025
Current Large Language Models (LLMs) exhibit limited ability to understand table structures and to apply precise numerical reasoning, which is crucial for tasks such as table question answering (TQA) and table-based fact verification (TFV). To address these challenges, we introduce our Tool-Augmented Reasoning framework for Tables (TART), which integrates LLMs with specialized tools. TART contains three key components: a table formatter to ensure accurate data representation, a tool maker to develop specific computational tools, and an explanation generator to maintain explainability. We also present the TOOLTAB dataset, a new benchmark designed specifically for training LLMs in table-tool integration. Our experiments indicate that TART achieves substantial improvements over existing methods (e.g., Chain-of-Thought) by improving both the precision of data processing and the clarity of the reasoning process. Notably, TART paired with CodeLlama achieves 90.0% of the accuracy of the closed-sourced LLM GPT-3.5-turbo, highlighting its robustness in diverse real-world scenarios. All the code and data are available at https://github.com/XinyuanLu00/TART.
comment: NAACL 2025 (Findings)
♻ ☆ CoAM: Corpus of All-Type Multiword Expressions ACL 2025
Multiword expressions (MWEs) refer to idiomatic sequences of multiple words. MWE identification, i.e., detecting MWEs in text, can play a key role in downstream tasks such as machine translation, but existing datasets for the task are inconsistently annotated, limited to a single type of MWE, or limited in size. To enable reliable and comprehensive evaluation, we created CoAM: Corpus of All-Type Multiword Expressions, a dataset of 1.3K sentences constructed through a multi-step process to enhance data quality consisting of human annotation, human review, and automated consistency checking. Additionally, for the first time in a dataset of MWE identification, CoAM's MWEs are tagged with MWE types, such as Noun and Verb, enabling fine-grained error analysis. Annotations for CoAM were collected using a new interface created with our interface generator, which allows easy and flexible annotation of MWEs in any form. Through experiments using CoAM, we find that a fine-tuned large language model outperforms MWEasWSD, which achieved the state-of-the-art performance on the DiMSUM dataset. Furthermore, analysis using our MWE type tagged data reveals that Verb MWEs are easier than Noun MWEs to identify across approaches.
comment: ACL 2025 main
♻ ☆ Rethinking Verification for LLM Code Generation: From Generation to Testing
Large language models (LLMs) have recently achieved notable success in code-generation benchmarks such as HumanEval and LiveCodeBench. However, a detailed examination reveals that these evaluation suites often comprise only a limited number of homogeneous test cases, resulting in subtle faults going undetected. This not only artificially inflates measured performance but also compromises accurate reward estimation in reinforcement learning frameworks utilizing verifiable rewards (RLVR). To address these critical shortcomings, we systematically investigate the test-case generation (TCG) task by proposing multi-dimensional metrics designed to rigorously quantify test-suite thoroughness. Furthermore, we introduce a human-LLM collaborative method (SAGA), leveraging human programming expertise with LLM reasoning capability, aimed at significantly enhancing both the coverage and the quality of generated test cases. In addition, we develop a TCGBench to facilitate the study of the TCG task. Experiments show that SAGA achieves a detection rate of 90.62% and a verifier accuracy of 32.58% on TCGBench. The Verifier Accuracy (Verifier Acc) of the code generation evaluation benchmark synthesized by SAGA is 10.78% higher than that of LiveCodeBench-v6. These results demonstrate the effectiveness of our proposed method. We hope this work contributes to building a scalable foundation for reliable LLM code evaluation, further advancing RLVR in code generation, and paving the way for automated adversarial test synthesis and adaptive benchmark integration.
♻ ☆ Large Language Model for Extracting Complex Contract Information in Industrial Scenes
This paper proposes a high-quality dataset construction method for complex contract information extraction tasks in industrial scenarios and fine-tunes a large language model based on this dataset. Firstly, cluster analysis is performed on industrial contract texts, and GPT-4 and GPT-3.5 are used to extract key information from the original contract data, obtaining high-quality data annotations. Secondly, data augmentation is achieved by constructing new texts, and GPT-3.5 generates unstructured contract texts from randomly combined keywords, improving model robustness. Finally, the large language model is fine-tuned based on the high-quality dataset. Experimental results show that the model achieves excellent overall performance while ensuring high field recall and precision and considering parsing efficiency. LoRA, data balancing, and data augmentation effectively enhance model accuracy and robustness. The proposed method provides a novel and efficient solution for industrial contract information extraction tasks.
♻ ☆ BountyBench: Dollar Impact of AI Agent Attackers and Defenders on Real-World Cybersecurity Systems
AI agents have the potential to significantly alter the cybersecurity landscape. Here, we introduce the first framework to capture offensive and defensive cyber-capabilities in evolving real-world systems. Instantiating this framework with BountyBench, we set up 25 systems with complex, real-world codebases. To capture the vulnerability lifecycle, we define three task types: Detect (detecting a new vulnerability), Exploit (exploiting a specific vulnerability), and Patch (patching a specific vulnerability). For Detect, we construct a new success indicator, which is general across vulnerability types and provides localized evaluation. We manually set up the environment for each system, including installing packages, setting up server(s), and hydrating database(s). We add 40 bug bounties, which are vulnerabilities with monetary awards of \$10-\$30,485, covering 9 of the OWASP Top 10 Risks. To modulate task difficulty, we devise a new strategy based on information to guide detection, interpolating from identifying a zero day to exploiting a specific vulnerability. We evaluate 8 agents: Claude Code, OpenAI Codex CLI with o3-high and o4-mini, and custom agents with o3-high, GPT-4.1, Gemini 2.5 Pro Preview, Claude 3.7 Sonnet Thinking, and DeepSeek-R1. Given up to three attempts, the top-performing agents are OpenAI Codex CLI: o3-high (12.5% on Detect, mapping to \$3,720; 90% on Patch, mapping to \$14,152), Custom Agent with Claude 3.7 Sonnet Thinking (67.5% on Exploit), and OpenAI Codex CLI: o4-mini (90% on Patch, mapping to \$14,422). OpenAI Codex CLI: o3-high, OpenAI Codex CLI: o4-mini, and Claude Code are more capable at defense, achieving higher Patch scores of 90%, 90%, and 87.5%, compared to Exploit scores of 47.5%, 32.5%, and 57.5% respectively; while the custom agents are relatively balanced between offense and defense, achieving Exploit scores of 37.5-67.5% and Patch scores of 35-60%.
comment: 93 pages
♻ ☆ Shifting from Ranking to Set Selection for Retrieval Augmented Generation ACL 2025
Retrieval in Retrieval-Augmented Generation(RAG) must ensure that retrieved passages are not only individually relevant but also collectively form a comprehensive set. Existing approaches primarily rerank top-k passages based on their individual relevance, often failing to meet the information needs of complex queries in multi-hop question answering. In this work, we propose a set-wise passage selection approach and introduce SETR, which explicitly identifies the information requirements of a query through Chain-of-Thought reasoning and selects an optimal set of passages that collectively satisfy those requirements. Experiments on multi-hop RAG benchmarks show that SETR outperforms both proprietary LLM-based rerankers and open-source baselines in terms of answer correctness and retrieval quality, providing an effective and efficient alternative to traditional rerankers in RAG systems. The code is available at https://github.com/LGAI-Research/SetR
comment: Accepted to ACL 2025 main (Oral Presentation)
♻ ☆ Post-hoc Study of Climate Microtargeting on Social Media Ads with LLMs: Thematic Insights and Fairness Evaluation
Climate change communication on social media increasingly employs microtargeting strategies to effectively reach and influence specific demographic groups. This study presents a post-hoc analysis of microtargeting practices within climate campaigns by leveraging large language models (LLMs) to examine Facebook advertisements. Our analysis focuses on two key aspects: demographic targeting and fairness. We evaluate the ability of LLMs to accurately predict the intended demographic targets, such as gender and age group, achieving an overall accuracy of 88.55%. Furthermore, we instruct the LLMs to generate explanations for their classifications, providing transparent reasoning behind each decision. These explanations reveal the specific thematic elements used to engage different demographic segments, highlighting distinct strategies tailored to various audiences. Our findings show that young adults are primarily targeted through messages emphasizing activism and environmental consciousness, while women are engaged through themes related to caregiving roles and social advocacy. In addition to evaluating the effectiveness of LLMs in detecting microtargeted messaging, we conduct a comprehensive fairness analysis to identify potential biases in model predictions. Our findings indicate that while LLMs perform well overall, certain biases exist, particularly in the classification of senior citizens and male audiences. By showcasing the efficacy of LLMs in dissecting and explaining targeted communication strategies and by highlighting fairness concerns, this study provides a valuable framework for future research aimed at enhancing transparency, accountability, and inclusivity in social media-driven climate campaigns.
♻ ☆ Extracting memorized pieces of (copyrighted) books from open-weight language models
Plaintiffs and defendants in copyright lawsuits over generative AI often make sweeping, opposing claims about the extent to which large language models (LLMs) have memorized plaintiffs' protected expression. Drawing on adversarial ML and copyright law, we show that these polarized positions dramatically oversimplify the relationship between memorization and copyright. To do so, we leverage a recent probabilistic extraction technique to extract pieces of the Books3 dataset from 17 open-weight LLMs. Through numerous experiments, we show that it's possible to extract substantial parts of at least some books from different LLMs. This is evidence that these LLMs have memorized the extracted text; this memorized content is copied inside the model parameters. But the results are complicated: the extent of memorization varies both by model and by book. With our specific experiments, we find that the largest LLMs don't memorize most books--either in whole or in part. However, we also find that Llama 3.1 70B memorizes some books, like Harry Potter and the Sorcerer's Stone and 1984, almost entirely. In fact, Harry Potter is so memorized that, using a seed prompt consisting of just the first line of chapter 1, we can deterministically generate the entire book near-verbatim. We discuss why our results have significant implications for copyright cases, though not ones that unambiguously favor either side.
♻ ☆ Riddle Generation using Learning Resources
One of the primary challenges in online learning environments, is to retain learner engagement. Several different instructional strategies are proposed both in online and offline environments to enhance learner engagement. The Concept Attainment Model is one such instructional strategy that focuses on learners acquiring a deeper understanding of a concept rather than just its dictionary definition. This is done by searching and listing the properties used to distinguish examples from non-examples of various concepts. Our work attempts to apply the Concept Attainment Model to build conceptual riddles, to deploy over online learning environments. The approach involves creating factual triples from learning resources, classifying them based on their uniqueness to a concept into `Topic Markers' and `Common', followed by generating riddles based on the Concept Attainment Model's format and capturing all possible solutions to those riddles. The results obtained from the human evaluation of riddles prove encouraging.
♻ ☆ GeistBERT: Breathing Life into German NLP
Advances in transformer-based language models have highlighted the benefits of language-specific pre-training on high-quality corpora. In this context, German NLP stands to gain from updated architectures and modern datasets tailored to the linguistic characteristics of the German language. GeistBERT seeks to improve German language processing by incrementally training on a diverse corpus and optimizing model performance across various NLP tasks. We pre-trained GeistBERT using fairseq, following the RoBERTa base configuration with Whole Word Masking (WWM), and initialized from GottBERT weights. The model was trained on a 1.3 TB German corpus with dynamic masking and a fixed sequence length of 512 tokens. For evaluation, we fine-tuned the model on standard downstream tasks, including NER (CoNLL 2003, GermEval 2014), text classification (GermEval 2018 coarse/fine, 10kGNAD), and NLI (German XNLI), using $F_1$ score and accuracy as evaluation metrics. GeistBERT achieved strong results across all tasks, leading among base models and setting a new state-of-the-art (SOTA) in GermEval 2018 fine text classification. It also outperformed several larger models, particularly in classification benchmarks. To support research in German NLP, we release GeistBERT under the MIT license.
♻ ☆ The Flaws of Others: An LLM-driven Framework for Scientific Knowledge Production
Large-language models turn writing into a live exchange between humans and software. We capture this new medium with a discursive-network model that treats people and LLMs as equal nodes and tracks how their statements circulate. Broadening the focus from isolated hallucinations, we define invalidation (any factual, logical, or structural breach) and show it follows four hazards: drift from truth, self-repair, fresh fabrication, and external detection. A general mathematical model of discursive networks is developed to provide valuable insights: A network governed only by drift and self-repair stabilizes at a modest error rate; adding fabrication reproduces the high rates seen in current LLMs. Giving each false claim even a small chance of peer review shifts the system to a truth-dominant state. We operationalize peer review with the open-source \emph{Flaws-of-Others (FOO) algorithm}: a configurable loop in which any set of agents critique one another while a harmoniser merges their verdicts. The takeaway is practical and cultural: reliability in this new medium comes not from perfecting single models but from wiring imperfect ones into networks that keep each other honest.
comment: 27 pages, 3 figures, 4 tables, 1 algorithm, 48 references
Computer Vision and Pattern Recognition 144
☆ Impact of Pretraining Word Co-occurrence on Compositional Generalization in Multimodal Models
CLIP and large multimodal models (LMMs) have better accuracy on examples involving concepts that are highly represented in the training data. However, the role of concept combinations in the training data on compositional generalization is largely unclear -- for instance, how does accuracy vary when a common object appears in an uncommon pairing with another object? In this paper, we investigate how word co-occurrence statistics in the pretraining dataset (a proxy for co-occurrence of visual concepts) impacts CLIP/LMM performance. To disentangle the effects of word co-occurrence frequencies from single-word frequencies, we measure co-occurrence with pointwise mutual information (PMI), which normalizes the joint probability of two words co-occurring by the probability of co-occurring independently. Using synthetically generated images with a variety of concept pairs, we show a strong correlation between PMI in the CLIP pretraining data and zero-shot accuracy in CLIP models trained on LAION-400M (r=0.97 and 14% accuracy gap between images in the top and bottom 5% of PMI values), demonstrating that even accuracy on common concepts is affected by the combination of concepts in the image. Leveraging this finding, we reproduce this effect in natural images by editing them to contain pairs with varying PMI, resulting in a correlation of r=0.75. Finally, we demonstrate that this behavior in CLIP transfers to LMMs built on top of CLIP (r=0.70 for TextVQA, r=0.62 for VQAv2). Our findings highlight the need for algorithms and architectures that improve compositional generalization in multimodal models without scaling the training data combinatorially. Our code is available at https://github.com/helenqu/multimodal-pretraining-pmi.
☆ Traceable Evidence Enhanced Visual Grounded Reasoning: Evaluation and Methodology
Models like OpenAI-o3 pioneer visual grounded reasoning by dynamically referencing visual regions, just like human "thinking with images". However, no benchmark exists to evaluate these capabilities holistically. To bridge this gap, we propose TreeBench (Traceable Evidence Evaluation Benchmark), a diagnostic benchmark built on three principles: (1) focused visual perception of subtle targets in complex scenes, (2) traceable evidence via bounding box evaluation, and (3) second-order reasoning to test object interactions and spatial hierarchies beyond simple object localization. Prioritizing images with dense objects, we initially sample 1K high-quality images from SA-1B, and incorporate eight LMM experts to manually annotate questions, candidate options, and answers for each image. After three stages of quality control, TreeBench consists of 405 challenging visual question-answering pairs, even the most advanced models struggle with this benchmark, where none of them reach 60% accuracy, e.g., OpenAI-o3 scores only 54.87. Furthermore, we introduce TreeVGR (Traceable Evidence Enhanced Visual Grounded Reasoning), a training paradigm to supervise localization and reasoning jointly with reinforcement learning, enabling accurate localizations and explainable reasoning pathways. Initialized from Qwen2.5-VL-7B, it improves V* Bench (+16.8), MME-RealWorld (+12.6), and TreeBench (+13.4), proving traceability is key to advancing vision-grounded reasoning. The code is available at https://github.com/Haochen-Wang409/TreeVGR.
☆ PyVision: Agentic Vision with Dynamic Tooling
LLMs are increasingly deployed as agents, systems capable of planning, reasoning, and dynamically calling external tools. However, in visual reasoning, prior approaches largely remain limited by predefined workflows and static toolsets. In this report, we present PyVision, an interactive, multi-turn framework that enables MLLMs to autonomously generate, execute, and refine Python-based tools tailored to the task at hand, unlocking flexible and interpretable problem-solving. We develop a taxonomy of the tools created by PyVision and analyze their usage across a diverse set of benchmarks. Quantitatively, PyVision achieves consistent performance gains, boosting GPT-4.1 by +7.8% on V* and Claude-4.0-Sonnet by +31.1% on VLMsAreBlind-mini. These results point to a broader shift: dynamic tooling allows models not just to use tools, but to invent them, advancing toward more agentic visual reasoning.
comment: 26 Pages, 10 Figures, Technical report
☆ MGVQ: Could VQ-VAE Beat VAE? A Generalizable Tokenizer with Multi-group Quantization
Vector Quantized Variational Autoencoders (VQ-VAEs) are fundamental models that compress continuous visual data into discrete tokens. Existing methods have tried to improve the quantization strategy for better reconstruction quality, however, there still exists a large gap between VQ-VAEs and VAEs. To narrow this gap, we propose \NickName, a novel method to augment the representation capability of discrete codebooks, facilitating easier optimization for codebooks and minimizing information loss, thereby enhancing reconstruction quality. Specifically, we propose to retain the latent dimension to preserve encoded features and incorporate a set of sub-codebooks for quantization. Furthermore, we construct comprehensive zero-shot benchmarks featuring resolutions of 512p and 2k to evaluate the reconstruction performance of existing methods rigorously. \NickName~achieves the \textbf{state-of-the-art performance on both ImageNet and $8$ zero-shot benchmarks} across all VQ-VAEs. Notably, compared with SD-VAE, we outperform them on ImageNet significantly, with rFID $\textbf{0.49}$ v.s. $\textbf{0.91}$, and achieve superior PSNR on all zero-shot benchmarks. These results highlight the superiority of \NickName~in reconstruction and pave the way for preserving fidelity in HD image processing tasks. Code will be publicly available at https://github.com/MKJia/MGVQ.
☆ Single-pass Adaptive Image Tokenization for Minimum Program Search
According to Algorithmic Information Theory (AIT) -- Intelligent representations compress data into the shortest possible program that can reconstruct its content, exhibiting low Kolmogorov Complexity (KC). In contrast, most visual representation learning systems use fixed-length representations for all inputs, ignoring variations in complexity or familiarity. Recent adaptive tokenization methods address this by allocating variable-length representations but typically require test-time search over multiple encodings to find the most predictive one. Inspired by Kolmogorov Complexity principles, we propose a single-pass adaptive tokenizer, KARL, which predicts the appropriate number of tokens for an image in a single forward pass, halting once its approximate KC is reached. The token count serves as a proxy for the minimum description length. KARL's training procedure closely resembles the Upside-Down Reinforcement Learning paradigm, as it learns to conditionally predict token halting based on a desired reconstruction quality. KARL matches the performance of recent adaptive tokenizers while operating in a single pass. We present scaling laws for KARL, analyzing the role of encoder/decoder size, continuous vs. discrete tokenization and more. Additionally, we offer a conceptual study drawing an analogy between Adaptive Image Tokenization and Algorithmic Information Theory, examining the predicted image complexity (KC) across axes such as structure vs. noise and in- vs. out-of-distribution familiarity -- revealing alignment with human intuition.
comment: Code at: https://github.com/ShivamDuggal4/karl Keywords: Representation Learning, Adaptive Tokenization, Compression, Algorithmic Information Theory, Kolmogorov Complexity, Upside-Down RL
☆ Doodle Your Keypoints: Sketch-Based Few-Shot Keypoint Detection ICCV 2025
Keypoint detection, integral to modern machine perception, faces challenges in few-shot learning, particularly when source data from the same distribution as the query is unavailable. This gap is addressed by leveraging sketches, a popular form of human expression, providing a source-free alternative. However, challenges arise in mastering cross-modal embeddings and handling user-specific sketch styles. Our proposed framework overcomes these hurdles with a prototypical setup, combined with a grid-based locator and prototypical domain adaptation. We also demonstrate success in few-shot convergence across novel keypoints and classes through extensive experiments.
comment: Accepted at ICCV 2025. Project Page: https://subhajitmaity.me/DYKp
☆ Multigranular Evaluation for Brain Visual Decoding
Existing evaluation protocols for brain visual decoding predominantly rely on coarse metrics that obscure inter-model differences, lack neuroscientific foundation, and fail to capture fine-grained visual distinctions. To address these limitations, we introduce BASIC, a unified, multigranular evaluation framework that jointly quantifies structural fidelity, inferential alignment, and contextual coherence between decoded and ground truth images. For the structural level, we introduce a hierarchical suite of segmentation-based metrics, including foreground, semantic, instance, and component masks, anchored in granularity-aware correspondence across mask structures. For the semantic level, we extract structured scene representations encompassing objects, attributes, and relationships using multimodal large language models, enabling detailed, scalable, and context-rich comparisons with ground-truth stimuli. We benchmark a diverse set of visual decoding methods across multiple stimulus-neuroimaging datasets within this unified evaluation framework. Together, these criteria provide a more discriminative, interpretable, and comprehensive foundation for measuring brain visual decoding methods.
comment: Project: https://weihaox.github.io/BASIC
☆ Multi-Granular Spatio-Temporal Token Merging for Training-Free Acceleration of Video LLMs ICCV2025
Video large language models (LLMs) achieve strong video understanding by leveraging a large number of spatio-temporal tokens, but suffer from quadratic computational scaling with token count. To address this, we propose a training-free spatio-temporal token merging method, named STTM. Our key insight is to exploit local spatial and temporal redundancy in video data which has been overlooked in prior work. STTM first transforms each frame into multi-granular spatial tokens using a coarse-to-fine search over a quadtree structure, then performs directed pairwise merging across the temporal dimension. This decomposed merging approach outperforms existing token reduction methods across six video QA benchmarks. Notably, STTM achieves a 2$\times$ speed-up with only a 0.5% accuracy drop under a 50% token budget, and a 3$\times$ speed-up with just a 2% drop under a 30% budget. Moreover, STTM is query-agnostic, allowing KV cache reuse across different questions for the same video. The project page is available at https://www.jshyun.me/projects/sttm.
comment: Accepted at ICCV2025; Project page: https://www.jshyun.me/projects/sttm
☆ CLIP Won't Learn Object-Attribute Binding from Natural Data and Here is Why
Contrastive vision-language models like CLIP are used for a large variety of applications, such as zero-shot classification or as vision encoder for multi-modal models. Despite their popularity, their representations show major limitations. For instance, CLIP models learn bag-of-words representations and, as a consequence, fail to distinguish whether an image is of "a yellow submarine and a blue bus" or "a blue submarine and a yellow bus". Previous attempts to fix this issue added hard negatives during training or modified the architecture, but failed to resolve the problem in its entirety. We suspect that the missing insights to solve the binding problem for CLIP are hidden in the arguably most important part of learning algorithms: the data. In this work, we fill this gap by rigorously identifying the influence of data properties on CLIP's ability to learn binding using a synthetic dataset. We find that common properties of natural data such as low attribute density, incomplete captions, and the saliency bias, a tendency of human captioners to describe the object that is "most salient" to them have a detrimental effect on binding performance. In contrast to common belief, we find that neither scaling the batch size, i.e., implicitly adding more hard negatives, nor explicitly creating hard negatives enables CLIP to learn reliable binding. Only when the data expresses our identified data properties CLIP learns almost perfect binding.
☆ OST-Bench: Evaluating the Capabilities of MLLMs in Online Spatio-temporal Scene Understanding
Recent advances in multimodal large language models (MLLMs) have shown remarkable capabilities in integrating vision and language for complex reasoning. While most existing benchmarks evaluate models under offline settings with a fixed set of pre-recorded inputs, we introduce OST-Bench, a benchmark designed to evaluate Online Spatio-Temporal understanding from the perspective of an agent actively exploring a scene. The Online aspect emphasizes the need to process and reason over incrementally acquired observations, while the Spatio-Temporal component requires integrating current visual inputs with historical memory to support dynamic spatial reasoning. OST-Bench better reflects the challenges of real-world embodied perception. Built on an efficient data collection pipeline, OST-Bench consists of 1.4k scenes and 10k question-answer pairs collected from ScanNet, Matterport3D, and ARKitScenes. We evaluate several leading MLLMs on OST-Bench and observe that they fall short on tasks requiring complex spatio-temporal reasoning. Under the online setting, their accuracy declines as the exploration horizon extends and the memory grows. Through further experimental analysis, we identify common error patterns across models and find that both complex clue-based spatial reasoning demands and long-term memory retrieval requirements significantly drop model performance along two separate axes, highlighting the core challenges that must be addressed to improve online embodied reasoning. To foster further research and development in the field, our codes, dataset, and benchmark are available. Our project page is: https://rbler1234.github.io/OSTBench.github.io/
comment: 28 pages, a benchmark designed to evaluate Online Spatio-Temporal understanding from the perspective of an agent actively exploring a scene. Project Page: https://rbler1234.github.io/OSTBench.github.io/
☆ Geometry Forcing: Marrying Video Diffusion and 3D Representation for Consistent World Modeling
Videos inherently represent 2D projections of a dynamic 3D world. However, our analysis suggests that video diffusion models trained solely on raw video data often fail to capture meaningful geometric-aware structure in their learned representations. To bridge this gap between video diffusion models and the underlying 3D nature of the physical world, we propose Geometry Forcing, a simple yet effective method that encourages video diffusion models to internalize latent 3D representations. Our key insight is to guide the model's intermediate representations toward geometry-aware structure by aligning them with features from a pretrained geometric foundation model. To this end, we introduce two complementary alignment objectives: Angular Alignment, which enforces directional consistency via cosine similarity, and Scale Alignment, which preserves scale-related information by regressing unnormalized geometric features from normalized diffusion representation. We evaluate Geometry Forcing on both camera view-conditioned and action-conditioned video generation tasks. Experimental results demonstrate that our method substantially improves visual quality and 3D consistency over the baseline methods. Project page: https://GeometryForcing.github.io.
comment: 18 pages, project page: https://GeometryForcing.github.io
☆ Martian World Models: Controllable Video Synthesis with Physically Accurate 3D Reconstructions
Synthesizing realistic Martian landscape videos is crucial for mission rehearsal and robotic simulation. However, this task poses unique challenges due to the scarcity of high-quality Martian data and the significant domain gap between Martian and terrestrial imagery. To address these challenges, we propose a holistic solution composed of two key components: 1) A data curation pipeline Multimodal Mars Synthesis (M3arsSynth), which reconstructs 3D Martian environments from real stereo navigation images, sourced from NASA's Planetary Data System (PDS), and renders high-fidelity multiview 3D video sequences. 2) A Martian terrain video generator, MarsGen, which synthesizes novel videos visually realistic and geometrically consistent with the 3D structure encoded in the data. Our M3arsSynth engine spans a wide range of Martian terrains and acquisition dates, enabling the generation of physically accurate 3D surface models at metric-scale resolution. MarsGen, fine-tuned on M3arsSynth data, synthesizes videos conditioned on an initial image frame and, optionally, camera trajectories or textual prompts, allowing for video generation in novel environments. Experimental results show that our approach outperforms video synthesis models trained on terrestrial datasets, achieving superior visual fidelity and 3D structural consistency.
comment: Project Page: https://marsgenai.github.io
☆ Scaling RL to Long Videos
We introduce a full-stack framework that scales up reasoning in vision-language models (VLMs) to long videos, leveraging reinforcement learning. We address the unique challenges of long video reasoning by integrating three critical components: (1) a large-scale dataset, LongVideo-Reason, comprising 52K long video QA pairs with high-quality reasoning annotations across diverse domains such as sports, games, and vlogs; (2) a two-stage training pipeline that extends VLMs with chain-of-thought supervised fine-tuning (CoT-SFT) and reinforcement learning (RL); and (3) a training infrastructure for long video RL, named Multi-modal Reinforcement Sequence Parallelism (MR-SP), which incorporates sequence parallelism and a vLLM-based engine tailored for long video, using cached video embeddings for efficient rollout and prefilling. In experiments, LongVILA-R1-7B achieves strong performance on long video QA benchmarks such as VideoMME. It also outperforms Video-R1-7B and even matches Gemini-1.5-Pro across temporal reasoning, goal and purpose reasoning, spatial reasoning, and plot reasoning on our LongVideo-Reason-eval benchmark. Notably, our MR-SP system achieves up to 2.1x speedup on long video RL training. LongVILA-R1 demonstrates consistent performance gains as the number of input video frames scales. LongVILA-R1 marks a firm step towards long video reasoning in VLMs. In addition, we release our training system for public availability that supports RL training on various modalities (video, text, and audio), various models (VILA and Qwen series), and even image and video generation models. On a single A100 node (8 GPUs), it supports RL training on hour-long videos (e.g., 3,600 frames / around 256k tokens).
comment: Code and models are available at https://github.com/NVlabs/Long-RL
☆ Input Conditioned Layer Dropping in Speech Foundation Models SP 2025
Curating foundation speech models for edge and IoT settings, where computational resources vary over time, requires dynamic architectures featuring adaptable reduction strategies. One emerging approach is layer dropping ($\mathcal{LD}$) which skips fraction of the layers of a backbone network during inference to reduce the computational load. This allows transforming static models into dynamic ones. However, existing approaches exhibit limitations either in the mode of selecting layers or by significantly modifying the neural architecture. To this end, we propose input-driven $\mathcal{LD}$ that employs the network's input features and a lightweight layer selecting network to determine the optimum combination of processing layers. Extensive experimentation on 4 speech and audio public benchmarks, using two different pre-trained foundation models, demonstrates the effectiveness of our approach, thoroughly outperforming random dropping and producing on-par (or better) results to early exit.
comment: Accepted at IEEE MLSP 2025
☆ TinierHAR: Towards Ultra-Lightweight Deep Learning Models for Efficient Human Activity Recognition on Edge Devices
Human Activity Recognition (HAR) on resource-constrained wearable devices demands inference models that harmonize accuracy with computational efficiency. This paper introduces TinierHAR, an ultra-lightweight deep learning architecture that synergizes residual depthwise separable convolutions, gated recurrent units (GRUs), and temporal aggregation to achieve SOTA efficiency without compromising performance. Evaluated across 14 public HAR datasets, TinierHAR reduces Parameters by 2.7x (vs. TinyHAR) and 43.3x (vs. DeepConvLSTM), and MACs by 6.4x and 58.6x, respectively, while maintaining the averaged F1-scores. Beyond quantitative gains, this work provides the first systematic ablation study dissecting the contributions of spatial-temporal components across proposed TinierHAR, prior SOTA TinyHAR, and the classical DeepConvLSTM, offering actionable insights for designing efficient HAR systems. We finally discussed the findings and suggested principled design guidelines for future efficient HAR. To catalyze edge-HAR research, we open-source all materials in this work for future benchmarking\footnote{https://github.com/zhaxidele/TinierHAR}
☆ Towards Continuous Home Cage Monitoring: An Evaluation of Tracking and Identification Strategies for Laboratory Mice
Continuous, automated monitoring of laboratory mice enables more accurate data collection and improves animal welfare through real-time insights. Researchers can achieve a more dynamic and clinically relevant characterization of disease progression and therapeutic effects by integrating behavioral and physiological monitoring in the home cage. However, providing individual mouse metrics is difficult because of their housing density, similar appearances, high mobility, and frequent interactions. To address these challenges, we develop a real-time identification (ID) algorithm that accurately assigns ID predictions to mice wearing custom ear tags in digital home cages monitored by cameras. Our pipeline consists of three parts: (1) a custom multiple object tracker (MouseTracks) that combines appearance and motion cues from mice; (2) a transformer-based ID classifier (Mouseformer); and (3) a tracklet associator linear program to assign final ID predictions to tracklets (MouseMap). Our models assign an animal ID based on custom ear tags at 30 frames per second with 24/7 cage coverage. We show that our custom tracking and ID pipeline improves tracking efficiency and lowers ID switches across mouse strains and various environmental factors compared to current mouse tracking methods.
☆ ArteryX: Advancing Brain Artery Feature Extraction with Vessel-Fused Networks and a Robust Validation Framework
Cerebrovascular pathology significantly contributes to cognitive decline and neurological disorders, underscoring the need for advanced tools to assess vascular integrity. Three-dimensional Time-of-Flight Magnetic Resonance Angiography (3D TOF MRA) is widely used to visualize cerebral vasculature, however, clinical evaluations generally focus on major arterial abnormalities, overlooking quantitative metrics critical for understanding subtle vascular changes. Existing methods for extracting structural, geometrical and morphological arterial features from MRA - whether manual or automated - face challenges including user-dependent variability, steep learning curves, and lack of standardized quantitative validations. We propose a novel semi-supervised artery evaluation framework, named ArteryX, a MATLAB-based toolbox that quantifies vascular features with high accuracy and efficiency, achieving processing times ~10-15 minutes per subject at 0.5 mm resolution with minimal user intervention. ArteryX employs a vessel-fused network based landmarking approach to reliably track and manage tracings, effectively addressing the issue of dangling/disconnected vessels. Validation on human subjects with cerebral small vessel disease demonstrated its improved sensitivity to subtle vascular changes and better performance than an existing semi-automated method. Importantly, the ArteryX toolbox enables quantitative feature validation by integrating an in-vivo like artery simulation framework utilizing vessel-fused graph nodes and predefined ground-truth features for specific artery types. Thus, the ArteryX framework holds promise for benchmarking feature extraction toolboxes and for seamless integration into clinical workflows, enabling early detection of cerebrovascular pathology and standardized comparisons across patient cohorts to advance understanding of vascular contributions to brain health.
comment: 14 Pages, 8 Figures, Preliminary version of the toolbox was presented at the ISMRM 2025 Conference in Hawaii at the "Software Tools" Session
☆ Not Only Consistency: Enhance Test-Time Adaptation with Spatio-temporal Inconsistency for Remote Physiological Measurement
Remote photoplethysmography (rPPG) has emerged as a promising non-invasive method for monitoring physiological signals using the camera. Although various domain adaptation and generalization methods were proposed to promote the adaptability of deep-based rPPG models in unseen deployment environments, considerations in aspects like privacy concerns and real-time adaptation restrict their application in real-world deployment. Thus, we aim to propose a novel fully Test-Time Adaptation (TTA) strategy tailored for rPPG tasks in this work. Specifically, based on prior knowledge in physiology and our observations, we noticed not only there is spatio-temporal consistency in the frequency domain of rPPG signals, but also that inconsistency in the time domain was significant. Given this, by leveraging both consistency and inconsistency priors, we introduce an innovative expert knowledge-based self-supervised \textbf{C}onsistency-\textbf{i}n\textbf{C}onsistency-\textbf{i}ntegration (\textbf{CiCi}) framework to enhances model adaptation during inference. Besides, our approach further incorporates a gradient dynamic control mechanism to mitigate potential conflicts between priors, ensuring stable adaptation across instances. Through extensive experiments on five diverse datasets under the TTA protocol, our method consistently outperforms existing techniques, presenting state-of-the-art performance in real-time self-supervised adaptation without accessing source data. The code will be released later.
☆ Hardware-Aware Feature Extraction Quantisation for Real-Time Visual Odometry on FPGA Platforms
Accurate position estimation is essential for modern navigation systems deployed in autonomous platforms, including ground vehicles, marine vessels, and aerial drones. In this context, Visual Simultaneous Localisation and Mapping (VSLAM) - which includes Visual Odometry - relies heavily on the reliable extraction of salient feature points from the visual input data. In this work, we propose an embedded implementation of an unsupervised architecture capable of detecting and describing feature points. It is based on a quantised SuperPoint convolutional neural network. Our objective is to minimise the computational demands of the model while preserving high detection quality, thus facilitating efficient deployment on platforms with limited resources, such as mobile or embedded systems. We implemented the solution on an FPGA System-on-Chip (SoC) platform, specifically the AMD/Xilinx Zynq UltraScale+, where we evaluated the performance of Deep Learning Processing Units (DPUs) and we also used the Brevitas library and the FINN framework to perform model quantisation and hardware-aware optimisation. This allowed us to process 640 x 480 pixel images at up to 54 fps on an FPGA platform, outperforming state-of-the-art solutions in the field. We conducted experiments on the TUM dataset to demonstrate and discuss the impact of different quantisation techniques on the accuracy and performance of the model in a visual odometry task.
comment: Accepted for the DSD 2025 conference in Salerno, Italy
☆ MIRA: A Novel Framework for Fusing Modalities in Medical RAG
Multimodal Large Language Models (MLLMs) have significantly advanced AI-assisted medical diagnosis, but they often generate factually inconsistent responses that deviate from established medical knowledge. Retrieval-Augmented Generation (RAG) enhances factual accuracy by integrating external sources, but it presents two key challenges. First, insufficient retrieval can miss critical information, whereas excessive retrieval can introduce irrelevant or misleading content, disrupting model output. Second, even when the model initially provides correct answers, over-reliance on retrieved data can lead to factual errors. To address these issues, we introduce the Multimodal Intelligent Retrieval and Augmentation (MIRA) framework, designed to optimize factual accuracy in MLLM. MIRA consists of two key components: (1) a calibrated Rethinking and Rearrangement module that dynamically adjusts the number of retrieved contexts to manage factual risk, and (2) A medical RAG framework integrating image embeddings and a medical knowledge base with a query-rewrite module for efficient multimodal reasoning. This enables the model to effectively integrate both its inherent knowledge and external references. Our evaluation of publicly available medical VQA and report generation benchmarks demonstrates that MIRA substantially enhances factual accuracy and overall performance, achieving new state-of-the-art results. Code is released at https://github.com/mbzuai-oryx/MIRA.
comment: ACM Multimedia 2025
☆ Single-Step Latent Diffusion for Underwater Image Restoration
Underwater image restoration algorithms seek to restore the color, contrast, and appearance of a scene that is imaged underwater. They are a critical tool in applications ranging from marine ecology and aquaculture to underwater construction and archaeology. While existing pixel-domain diffusion-based image restoration approaches are effective at restoring simple scenes with limited depth variation, they are computationally intensive and often generate unrealistic artifacts when applied to scenes with complex geometry and significant depth variation. In this work we overcome these limitations by combining a novel network architecture (SLURPP) with an accurate synthetic data generation pipeline. SLURPP combines pretrained latent diffusion models -- which encode strong priors on the geometry and depth of scenes -- with an explicit scene decomposition -- which allows one to model and account for the effects of light attenuation and backscattering. To train SLURPP we design a physics-based underwater image synthesis pipeline that applies varied and realistic underwater degradation effects to existing terrestrial image datasets. This approach enables the generation of diverse training data with dense medium/degradation annotations. We evaluate our method extensively on both synthetic and real-world benchmarks and demonstrate state-of-the-art performance. Notably, SLURPP is over 200X faster than existing diffusion-based methods while offering ~ 3 dB improvement in PSNR on synthetic benchmarks. It also offers compelling qualitative improvements on real-world data. Project website https://tianfwang.github.io/slurpp/.
☆ THUNDER: Tile-level Histopathology image UNDERstanding benchmark
Progress in a research field can be hard to assess, in particular when many concurrent methods are proposed in a short period of time. This is the case in digital pathology, where many foundation models have been released recently to serve as feature extractors for tile-level images, being used in a variety of downstream tasks, both for tile- and slide-level problems. Benchmarking available methods then becomes paramount to get a clearer view of the research landscape. In particular, in critical domains such as healthcare, a benchmark should not only focus on evaluating downstream performance, but also provide insights about the main differences between methods, and importantly, further consider uncertainty and robustness to ensure a reliable usage of proposed models. For these reasons, we introduce THUNDER, a tile-level benchmark for digital pathology foundation models, allowing for efficient comparison of many models on diverse datasets with a series of downstream tasks, studying their feature spaces and assessing the robustness and uncertainty of predictions informed by their embeddings. THUNDER is a fast, easy-to-use, dynamic benchmark that can already support a large variety of state-of-the-art foundation, as well as local user-defined models for direct tile-based comparison. In this paper, we provide a comprehensive comparison of 23 foundation models on 16 different datasets covering diverse tasks, feature analysis, and robustness. The code for THUNDER is publicly available at https://github.com/MICS-Lab/thunder.
☆ MeD-3D: A Multimodal Deep Learning Framework for Precise Recurrence Prediction in Clear Cell Renal Cell Carcinoma (ccRCC)
Accurate prediction of recurrence in clear cell renal cell carcinoma (ccRCC) remains a major clinical challenge due to the disease complex molecular, pathological, and clinical heterogeneity. Traditional prognostic models, which rely on single data modalities such as radiology, histopathology, or genomics, often fail to capture the full spectrum of disease complexity, resulting in suboptimal predictive accuracy. This study aims to overcome these limitations by proposing a deep learning (DL) framework that integrates multimodal data, including CT, MRI, histopathology whole slide images (WSI), clinical data, and genomic profiles, to improve the prediction of ccRCC recurrence and enhance clinical decision-making. The proposed framework utilizes a comprehensive dataset curated from multiple publicly available sources, including TCGA, TCIA, and CPTAC. To process the diverse modalities, domain-specific models are employed: CLAM, a ResNet50-based model, is used for histopathology WSIs, while MeD-3D, a pre-trained 3D-ResNet18 model, processes CT and MRI images. For structured clinical and genomic data, a multi-layer perceptron (MLP) is used. These models are designed to extract deep feature embeddings from each modality, which are then fused through an early and late integration architecture. This fusion strategy enables the model to combine complementary information from multiple sources. Additionally, the framework is designed to handle incomplete data, a common challenge in clinical settings, by enabling inference even when certain modalities are missing.
☆ 3D-ADAM: A Dataset for 3D Anomaly Detection in Advanced Manufacturing
Surface defects are one of the largest contributors to low yield in the manufacturing sector. Accurate and reliable detection of defects during the manufacturing process is therefore of great value across the sector. State-of-the-art approaches to automated defect detection yield impressive performance on current datasets, yet still fall short in real-world manufacturing settings and developing improved methods relies on large datasets representative of real-world scenarios. Unfortunately, high-quality, high-precision RGB+3D industrial anomaly detection datasets are scarce, and typically do not reflect real-world industrial deployment scenarios. To address this, we introduce 3D-ADAM, the first large-scale industry-relevant dataset for high-precision 3D Anomaly Detection. 3D-ADAM comprises 14,120 high-resolution scans across 217 unique parts, captured using 4 industrial depth imaging sensors. It includes 27,346 annotated defect instances from 12 categories, covering the breadth of industrial surface defects. 3D-ADAM uniquely captures an additional 8,110 annotations of machine element features, spanning the range of relevant mechanical design form factors. Unlike existing datasets, 3D-ADAM is captured in a real industrial environment with variations in part position and orientation, camera positioning, ambient lighting conditions, as well as partial occlusions. Our evaluation of SOTA models across various RGB+3D anomaly detection tasks demonstrates the significant challenge this dataset presents to current approaches. We further validated the industrial relevance and quality of the dataset through an expert labelling survey conducted by industry partners. By providing this challenging benchmark, 3D-ADAM aims to accelerate the development of robust 3D Anomaly Detection models capable of meeting the demands of modern manufacturing environments.
☆ Rethinking Query-based Transformer for Continual Image Segmentation CVPR 2025
Class-incremental/Continual image segmentation (CIS) aims to train an image segmenter in stages, where the set of available categories differs at each stage. To leverage the built-in objectness of query-based transformers, which mitigates catastrophic forgetting of mask proposals, current methods often decouple mask generation from the continual learning process. This study, however, identifies two key issues with decoupled frameworks: loss of plasticity and heavy reliance on input data order. To address these, we conduct an in-depth investigation of the built-in objectness and find that highly aggregated image features provide a shortcut for queries to generate masks through simple feature alignment. Based on this, we propose SimCIS, a simple yet powerful baseline for CIS. Its core idea is to directly select image features for query assignment, ensuring "perfect alignment" to preserve objectness, while simultaneously allowing queries to select new classes to promote plasticity. To further combat catastrophic forgetting of categories, we introduce cross-stage consistency in selection and an innovative "visual query"-based replay mechanism. Experiments demonstrate that SimCIS consistently outperforms state-of-the-art methods across various segmentation tasks, settings, splits, and input data orders. All models and codes will be made publicly available at https://github.com/SooLab/SimCIS.
comment: This work is accepted by CVPR 2025
☆ Benchmarking Content-Based Puzzle Solvers on Corrupted Jigsaw Puzzles
Content-based puzzle solvers have been extensively studied, demonstrating significant progress in computational techniques. However, their evaluation often lacks realistic challenges crucial for real-world applications, such as the reassembly of fragmented artefacts or shredded documents. In this work, we investigate the robustness of State-Of-The-Art content-based puzzle solvers introducing three types of jigsaw puzzle corruptions: missing pieces, eroded edges, and eroded contents. Evaluating both heuristic and deep learning-based solvers, we analyse their ability to handle these corruptions and identify key limitations. Our results show that solvers developed for standard puzzles have a rapid decline in performance if more pieces are corrupted. However, deep learning models can significantly improve their robustness through fine-tuning with augmented data. Notably, the advanced Positional Diffusion model adapts particularly well, outperforming its competitors in most experiments. Based on our findings, we highlight promising research directions for enhancing the automated reconstruction of real-world artefacts.
comment: Accepted at ICIAP 2025
☆ Patient-specific vs Multi-Patient Vision Transformer for Markerless Tumor Motion Forecasting
Background: Accurate forecasting of lung tumor motion is essential for precise dose delivery in proton therapy. While current markerless methods mostly rely on deep learning, transformer-based architectures remain unexplored in this domain, despite their proven performance in trajectory forecasting. Purpose: This work introduces a markerless forecasting approach for lung tumor motion using Vision Transformers (ViT). Two training strategies are evaluated under clinically realistic constraints: a patient-specific (PS) approach that learns individualized motion patterns, and a multi-patient (MP) model designed for generalization. The comparison explicitly accounts for the limited number of images that can be generated between planning and treatment sessions. Methods: Digitally reconstructed radiographs (DRRs) derived from planning 4DCT scans of 31 patients were used to train the MP model; a 32nd patient was held out for evaluation. PS models were trained using only the target patient's planning data. Both models used 16 DRRs per input and predicted tumor motion over a 1-second horizon. Performance was assessed using Average Displacement Error (ADE) and Final Displacement Error (FDE), on both planning (T1) and treatment (T2) data. Results: On T1 data, PS models outperformed MP models across all training set sizes, especially with larger datasets (up to 25,000 DRRs, p < 0.05). However, MP models demonstrated stronger robustness to inter-fractional anatomical variability and achieved comparable performance on T2 data without retraining. Conclusions: This is the first study to apply ViT architectures to markerless tumor motion forecasting. While PS models achieve higher precision, MP models offer robust out-of-the-box performance, well-suited for time-constrained clinical settings.
☆ Synergistic Prompting for Robust Visual Recognition with Missing Modalities
Large-scale multi-modal models have demonstrated remarkable performance across various visual recognition tasks by leveraging extensive paired multi-modal training data. However, in real-world applications, the presence of missing or incomplete modality inputs often leads to significant performance degradation. Recent research has focused on prompt-based strategies to tackle this issue; however, existing methods are hindered by two major limitations: (1) static prompts lack the flexibility to adapt to varying missing-data conditions, and (2) basic prompt-tuning methods struggle to ensure reliable performance when critical modalities are missing.To address these challenges, we propose a novel Synergistic Prompting (SyP) framework for robust visual recognition with missing modalities. The proposed SyP introduces two key innovations: (I) a Dynamic Adapter, which computes adaptive scaling factors to dynamically generate prompts, replacing static parameters for flexible multi-modal adaptation, and (II) a Synergistic Prompting Strategy, which combines static and dynamic prompts to balance information across modalities, ensuring robust reasoning even when key modalities are missing. The proposed SyP achieves significant performance improvements over existing approaches across three widely-used visual recognition datasets, demonstrating robustness under diverse missing rates and conditions. Extensive experiments and ablation studies validate its effectiveness in handling missing modalities, highlighting its superior adaptability and reliability.
☆ Adaptive Attention Residual U-Net for curvilinear structure segmentation in fluorescence microscopy and biomedical images
Segmenting curvilinear structures in fluorescence microscopy remains a challenging task, particularly under noisy conditions and in dense filament networks commonly seen in vivo. To address this, we created two original datasets consisting of hundreds of synthetic images of fluorescently labelled microtubules within cells. These datasets are precisely annotated and closely mimic real microscopy images, including realistic noise. The second dataset presents an additional challenge, by simulating varying fluorescence intensities along filaments that complicate segmentation. While deep learning has shown strong potential in biomedical image analysis, its performance often declines in noisy or low-contrast conditions. To overcome this limitation, we developed a novel advanced architecture: the Adaptive Squeeze-and-Excitation Residual U-Net (ASE_Res_UNet). This model enhanced the standard U-Net by integrating residual blocks in the encoder and adaptive SE attention mechanisms in the decoder. Through ablation studies and comprehensive visual and quantitative evaluations, ASE_Res_UNet consistently outperformed its variants, namely standard U-Net, ASE_UNet and Res_UNet architectures. These improvements, particularly in noise resilience and detecting fine, low-intensity structures, were largely attributed to the adaptive SE attention module that we created. We further benchmarked ASE_Res_UNet against various state-of-the-art models, and found it achieved superior performance on our most challenging dataset. Finally, the model also generalized well to real microscopy images of stained microtubules as well as to other curvilinear structures. Indeed, it successfully segmented retinal blood vessels and nerves in noisy or low-contrast biomedical images, demonstrating its strong potential for applications in disease diagnosis and treatment.
☆ Visual Instance-aware Prompt Tuning
Visual Prompt Tuning (VPT) has emerged as a parameter-efficient fine-tuning paradigm for vision transformers, with conventional approaches utilizing dataset-level prompts that remain the same across all input instances. We observe that this strategy results in sub-optimal performance due to high variance in downstream datasets. To address this challenge, we propose Visual Instance-aware Prompt Tuning (ViaPT), which generates instance-aware prompts based on each individual input and fuses them with dataset-level prompts, leveraging Principal Component Analysis (PCA) to retain important prompting information. Moreover, we reveal that VPT-Deep and VPT-Shallow represent two corner cases based on a conceptual understanding, in which they fail to effectively capture instance-specific information, while random dimension reduction on prompts only yields performance between the two extremes. Instead, ViaPT overcomes these limitations by balancing dataset-level and instance-level knowledge, while reducing the amount of learnable parameters compared to VPT-Deep. Extensive experiments across 34 diverse datasets demonstrate that our method consistently outperforms state-of-the-art baselines, establishing a new paradigm for analyzing and optimizing visual prompts for vision transformers.
☆ Robust and Generalizable Heart Rate Estimation via Deep Learning for Remote Photoplethysmography in Complex Scenarios
Non-contact remote photoplethysmography (rPPG) technology enables heart rate measurement from facial videos. However, existing network models still face challenges in accu racy, robustness, and generalization capability under complex scenarios. This paper proposes an end-to-end rPPG extraction network that employs 3D convolutional neural networks to reconstruct accurate rPPG signals from raw facial videos. We introduce a differential frame fusion module that integrates differential frames with original frames, enabling frame-level representations to capture blood volume pulse (BVP) variations. Additionally, we incorporate Temporal Shift Module (TSM) with self-attention mechanisms, which effectively enhance rPPG features with minimal computational overhead. Furthermore, we propose a novel dynamic hybrid loss function that provides stronger supervision for the network, effectively mitigating over fitting. Comprehensive experiments were conducted on not only the PURE and UBFC-rPPG datasets but also the challenging MMPD dataset under complex scenarios, involving both intra dataset and cross-dataset evaluations, which demonstrate the superior robustness and generalization capability of our network. Specifically, after training on PURE, our model achieved a mean absolute error (MAE) of 7.58 on the MMPD test set, outperforming the state-of-the-art models.
comment: 7 pages, 3 figures
☆ Computationally Efficient Information-Driven Optical Design with Interchanging Optimization
Recent work has demonstrated that imaging systems can be evaluated through the information content of their measurements alone, enabling application-agnostic optical design that avoids computational decoding challenges. Information-Driven Encoder Analysis Learning (IDEAL) was proposed to automate this process through gradient-based. In this work, we study IDEAL across diverse imaging systems and find that it suffers from high memory usage, long runtimes, and a potentially mismatched objective function due to end-to-end differentiability requirements. We introduce IDEAL with Interchanging Optimization (IDEAL-IO), a method that decouples density estimation from optical parameter optimization by alternating between fitting models to current measurements and updating optical parameters using fixed models for information estimation. This approach reduces runtime and memory usage by up to 6x while enabling more expressive density models that guide optimization toward superior designs. We validate our method on diffractive optics, lensless imaging, and snapshot 3D microscopy applications, establishing information-theoretic optimization as a practical, scalable strategy for real-world imaging system design.
☆ SURPRISE3D: A Dataset for Spatial Understanding and Reasoning in Complex 3D Scenes
The integration of language and 3D perception is critical for embodied AI and robotic systems to perceive, understand, and interact with the physical world. Spatial reasoning, a key capability for understanding spatial relationships between objects, remains underexplored in current 3D vision-language research. Existing datasets often mix semantic cues (e.g., object name) with spatial context, leading models to rely on superficial shortcuts rather than genuinely interpreting spatial relationships. To address this gap, we introduce S\textsc{urprise}3D, a novel dataset designed to evaluate language-guided spatial reasoning segmentation in complex 3D scenes. S\textsc{urprise}3D consists of more than 200k vision language pairs across 900+ detailed indoor scenes from ScanNet++ v2, including more than 2.8k unique object classes. The dataset contains 89k+ human-annotated spatial queries deliberately crafted without object name, thereby mitigating shortcut biases in spatial understanding. These queries comprehensively cover various spatial reasoning skills, such as relative position, narrative perspective, parametric perspective, and absolute distance reasoning. Initial benchmarks demonstrate significant challenges for current state-of-the-art expert 3D visual grounding methods and 3D-LLMs, underscoring the necessity of our dataset and the accompanying 3D Spatial Reasoning Segmentation (3D-SRS) benchmark suite. S\textsc{urprise}3D and 3D-SRS aim to facilitate advancements in spatially aware AI, paving the way for effective embodied interaction and robotic planning. The code and datasets can be found in https://github.com/liziwennba/SUPRISE.
☆ Where are we with calibration under dataset shift in image classification?
We conduct an extensive study on the state of calibration under real-world dataset shift for image classification. Our work provides important insights on the choice of post-hoc and in-training calibration techniques, and yields practical guidelines for all practitioners interested in robust calibration under shift. We compare various post-hoc calibration methods, and their interactions with common in-training calibration strategies (e.g., label smoothing), across a wide range of natural shifts, on eight different classification tasks across several imaging domains. We find that: (i) simultaneously applying entropy regularisation and label smoothing yield the best calibrated raw probabilities under dataset shift, (ii) post-hoc calibrators exposed to a small amount of semantic out-of-distribution data (unrelated to the task) are most robust under shift, (iii) recent calibration methods specifically aimed at increasing calibration under shifts do not necessarily offer significant improvements over simpler post-hoc calibration methods, (iv) improving calibration under shifts often comes at the cost of worsening in-distribution calibration. Importantly, these findings hold for randomly initialised classifiers, as well as for those finetuned from foundation models, the latter being consistently better calibrated compared to models trained from scratch. Finally, we conduct an in-depth analysis of ensembling effects, finding that (i) applying calibration prior to ensembling (instead of after) is more effective for calibration under shifts, (ii) for ensembles, OOD exposure deteriorates the ID-shifted calibration trade-off, (iii) ensembling remains one of the most effective methods to improve calibration robustness and, combined with finetuning from foundation models, yields best calibration results overall.
comment: Code available at https://github.com/biomedia-mira/calibration_under_shifts
☆ Synchronizing Task Behavior: Aligning Multiple Tasks during Test-Time Training ICCV 2025
Generalizing neural networks to unseen target domains is a significant challenge in real-world deployments. Test-time training (TTT) addresses this by using an auxiliary self-supervised task to reduce the domain gap caused by distribution shifts between the source and target. However, we find that when models are required to perform multiple tasks under domain shifts, conventional TTT methods suffer from unsynchronized task behavior, where the adaptation steps needed for optimal performance in one task may not align with the requirements of other tasks. To address this, we propose a novel TTT approach called Synchronizing Tasks for Test-time Training (S4T), which enables the concurrent handling of multiple tasks. The core idea behind S4T is that predicting task relations across domain shifts is key to synchronizing tasks during test time. To validate our approach, we apply S4T to conventional multi-task benchmarks, integrating it with traditional TTT protocols. Our empirical results show that S4T outperforms state-of-the-art TTT methods across various benchmarks.
comment: Accepted at ICCV 2025
☆ SCOOTER: A Human Evaluation Framework for Unrestricted Adversarial Examples
Unrestricted adversarial attacks aim to fool computer vision models without being constrained by $\ell_p$-norm bounds to remain imperceptible to humans, for example, by changing an object's color. This allows attackers to circumvent traditional, norm-bounded defense strategies such as adversarial training or certified defense strategies. However, due to their unrestricted nature, there are also no guarantees of norm-based imperceptibility, necessitating human evaluations to verify just how authentic these adversarial examples look. While some related work assesses this vital quality of adversarial attacks, none provide statistically significant insights. This issue necessitates a unified framework that supports and streamlines such an assessment for evaluating and comparing unrestricted attacks. To close this gap, we introduce SCOOTER - an open-source, statistically powered framework for evaluating unrestricted adversarial examples. Our contributions are: $(i)$ best-practice guidelines for crowd-study power, compensation, and Likert equivalence bounds to measure imperceptibility; $(ii)$ the first large-scale human vs. model comparison across 346 human participants showing that three color-space attacks and three diffusion-based attacks fail to produce imperceptible images. Furthermore, we found that GPT-4o can serve as a preliminary test for imperceptibility, but it only consistently detects adversarial examples for four out of six tested attacks; $(iii)$ open-source software tools, including a browser-based task template to collect annotations and analysis scripts in Python and R; $(iv)$ an ImageNet-derived benchmark dataset containing 3K real images, 7K adversarial examples, and over 34K human ratings. Our findings demonstrate that automated vision systems do not align with human perception, reinforcing the need for a ground-truth SCOOTER benchmark.
comment: 42 pages, 16 figures, 11 tables, Under Review, Code: https://github.com/DrenFazlija/Scooter, Data: https://doi.org/10.5281/zenodo.15771501
☆ Rainbow Artifacts from Electromagnetic Signal Injection Attacks on Image Sensors
Image sensors are integral to a wide range of safety- and security-critical systems, including surveillance infrastructure, autonomous vehicles, and industrial automation. These systems rely on the integrity of visual data to make decisions. In this work, we investigate a novel class of electromagnetic signal injection attacks that target the analog domain of image sensors, allowing adversaries to manipulate raw visual inputs without triggering conventional digital integrity checks. We uncover a previously undocumented attack phenomenon on CMOS image sensors: rainbow-like color artifacts induced in images captured by image sensors through carefully tuned electromagnetic interference. We further evaluate the impact of these attacks on state-of-the-art object detection models, showing that the injected artifacts propagate through the image signal processing pipeline and lead to significant mispredictions. Our findings highlight a critical and underexplored vulnerability in the visual perception stack, highlighting the need for more robust defenses against physical-layer attacks in such systems.
comment: 5 pages, 4 figures
☆ TRIX- Trading Adversarial Fairness via Mixed Adversarial Training
Adversarial Training (AT) is a widely adopted defense against adversarial examples. However, existing approaches typically apply a uniform training objective across all classes, overlooking disparities in class-wise vulnerability. This results in adversarial unfairness: classes with well distinguishable features (strong classes) tend to become more robust, while classes with overlapping or shared features(weak classes) remain disproportionately susceptible to adversarial attacks. We observe that strong classes do not require strong adversaries during training, as their non-robust features are quickly suppressed. In contrast, weak classes benefit from stronger adversaries to effectively reduce their vulnerabilities. Motivated by this, we introduce TRIX, a feature-aware adversarial training framework that adaptively assigns weaker targeted adversaries to strong classes, promoting feature diversity via uniformly sampled targets, and stronger untargeted adversaries to weak classes, enhancing their focused robustness. TRIX further incorporates per-class loss weighting and perturbation strength adjustments, building on prior work, to emphasize weak classes during the optimization. Comprehensive experiments on standard image classification benchmarks, including evaluations under strong attacks such as PGD and AutoAttack, demonstrate that TRIX significantly improves worst-case class accuracy on both clean and adversarial data, reducing inter-class robustness disparities, and preserves overall accuracy. Our results highlight TRIX as a practical step toward fair and effective adversarial defense.
☆ Deep Learning based 3D Volume Correlation for Additive Manufacturing Using High-Resolution Industrial X-ray Computed Tomography
Quality control in additive manufacturing (AM) is vital for industrial applications in areas such as the automotive, medical and aerospace sectors. Geometric inaccuracies caused by shrinkage and deformations can compromise the life and performance of additively manufactured components. Such deviations can be quantified using Digital Volume Correlation (DVC), which compares the computer-aided design (CAD) model with the X-ray Computed Tomography (XCT) geometry of the components produced. However, accurate registration between the two modalities is challenging due to the absence of a ground truth or reference deformation field. In addition, the extremely large data size of high-resolution XCT volumes makes computation difficult. In this work, we present a deep learning-based approach for estimating voxel-wise deformations between CAD and XCT volumes. Our method uses a dynamic patch-based processing strategy to handle high-resolution volumes. In addition to the Dice Score, we introduce a Binary Difference Map (BDM) that quantifies voxel-wise mismatches between binarized CAD and XCT volumes to evaluate the accuracy of the registration. Our approach shows a 9.2\% improvement in the Dice Score and a 9.9\% improvement in the voxel match rate compared to classic DVC methods, while reducing the interaction time from days to minutes. This work sets the foundation for deep learning-based DVC methods to generate compensation meshes that can then be used in closed-loop correlations during the AM production process. Such a system would be of great interest to industries since the manufacturing process will become more reliable and efficient, saving time and material.
☆ X-RAFT: Cross-Modal Non-Rigid Registration of Blue and White Light Neurosurgical Hyperspectral Images
Integration of hyperspectral imaging into fluorescence-guided neurosurgery has the potential to improve surgical decision making by providing quantitative fluorescence measurements in real-time. Quantitative fluorescence requires paired spectral data in fluorescence (blue light) and reflectance (white light) mode. Blue and white image acquisition needs to be performed sequentially in a potentially dynamic surgical environment. A key component to the fluorescence quantification process is therefore the ability to find dense cross-modal image correspondences between two hyperspectral images taken under these drastically different lighting conditions. We address this challenge with the introduction of X-RAFT, a Recurrent All-Pairs Field Transforms (RAFT) optical flow model modified for cross-modal inputs. We propose using distinct image encoders for each modality pair, and fine-tune these in a self-supervised manner using flow-cycle-consistency on our neurosurgical hyperspectral data. We show an error reduction of 36.6% across our evaluation metrics when comparing to a naive baseline and 27.83% reduction compared to an existing cross-modal optical flow method (CrossRAFT). Our code and models will be made publicly available after the review process.
☆ Sparse-Dense Side-Tuner for efficient Video Temporal Grounding
Video Temporal Grounding (VTG) involves Moment Retrieval (MR) and Highlight Detection (HD) based on textual queries. For this, most methods rely solely on final-layer features of frozen large pre-trained backbones, limiting their adaptability to new domains. While full fine-tuning is often impractical, parameter-efficient fine-tuning -- and particularly side-tuning (ST) -- has emerged as an effective alternative. However, prior ST approaches this problem from a frame-level refinement perspective, overlooking the inherent sparse nature of MR. To address this, we propose the Sparse-Dense Side-Tuner (SDST), the first anchor-free ST architecture for VTG. We also introduce the Reference-based Deformable Self-Attention, a novel mechanism that enhances the context modeling of the deformable attention -- a key limitation of existing anchor-free methods. Additionally, we present the first effective integration of InternVideo2 backbone into an ST framework, showing its profound implications in performance. Overall, our method significantly improves existing ST methods, achieving highly competitive or SOTA results on QVHighlights, TACoS, and Charades-STA, while reducing up to a 73% the parameter count w.r.t. the existing SOTA methods. The code is publicly accessible at https://github.com/davidpujol/SDST.
☆ EEvAct: Early Event-Based Action Recognition with High-Rate Two-Stream Spiking Neural Networks
Recognizing human activities early is crucial for the safety and responsiveness of human-robot and human-machine interfaces. Due to their high temporal resolution and low latency, event-based vision sensors are a perfect match for this early recognition demand. However, most existing processing approaches accumulate events to low-rate frames or space-time voxels which limits the early prediction capabilities. In contrast, spiking neural networks (SNNs) can process the events at a high-rate for early predictions, but most works still fall short on final accuracy. In this work, we introduce a high-rate two-stream SNN which closes this gap by outperforming previous work by 2% in final accuracy on the large-scale THU EACT-50 dataset. We benchmark the SNNs within a novel early event-based recognition framework by reporting Top-1 and Top-5 recognition scores for growing observation time. Finally, we exemplify the impact of these methods on a real-world task of early action triggering for human motion capture in sports.
comment: International Conference on Neuromorphic Systems (ICONS) 2025
☆ RTR-GS: 3D Gaussian Splatting for Inverse Rendering with Radiance Transfer and Reflection
3D Gaussian Splatting (3DGS) has demonstrated impressive capabilities in novel view synthesis. However, rendering reflective objects remains a significant challenge, particularly in inverse rendering and relighting. We introduce RTR-GS, a novel inverse rendering framework capable of robustly rendering objects with arbitrary reflectance properties, decomposing BRDF and lighting, and delivering credible relighting results. Given a collection of multi-view images, our method effectively recovers geometric structure through a hybrid rendering model that combines forward rendering for radiance transfer with deferred rendering for reflections. This approach successfully separates high-frequency and low-frequency appearances, mitigating floating artifacts caused by spherical harmonic overfitting when handling high-frequency details. We further refine BRDF and lighting decomposition using an additional physically-based deferred rendering branch. Experimental results show that our method enhances novel view synthesis, normal estimation, decomposition, and relighting while maintaining efficient training inference process.
comment: 16 pages
☆ Energy-Guided Decoding for Object Hallucination Mitigation
Mitigating object hallucination in large vision-language models (LVLMs) is critical to their safe deployment. Existing methods either are restricted to specific decoding methods, or demand sophisticated modifications to visual inputs, or rely on knowledge from external models. In this work, we first reveal the phenomenon that VLMs exhibit significant imbalance in the ``Yes'' ratio ( \ie, the fraction of ``Yes'' answers among the total number of questions) across three different visual question answering (VQA) datasets. Furthermore, we propose an energy-based decoding method, which dynamically selects the hidden states from the layer with minimal energy score. It is simple yet effective in reducing the bias for the yes ratio while boosting performance across three benchmarks (POPE, MME, and MMVP). Our method consistently improves accuracy and F1 score on three VQA datasets across three commonly used VLMs over several baseline methods. The average accuracy improvement is 4.82% compared to greedy decoding. Moreover, the average yes-ratio gap reduction is 8.81%, meaning the proposed method is less biased as shown in Figure 1.
☆ RAPS-3D: Efficient interactive segmentation for 3D radiological imaging
Promptable segmentation, introduced by the Segment Anything Model (SAM), is a promising approach for medical imaging, as it enables clinicians to guide and refine model predictions interactively. However, SAM's architecture is designed for 2D images and does not extend naturally to 3D volumetric data such as CT or MRI scans. Adapting 2D models to 3D typically involves autoregressive strategies, where predictions are propagated slice by slice, resulting in increased inference complexity. Processing large 3D volumes also requires significant computational resources, often leading existing 3D methods to also adopt complex strategies like sliding-window inference to manage memory usage, at the cost of longer inference times and greater implementation complexity. In this paper, we present a simplified 3D promptable segmentation method, inspired by SegVol, designed to reduce inference time and eliminate prompt management complexities associated with sliding windows while achieving state-of-the-art performance.
comment: Abstract accepted at MIUA 2025
☆ Understanding Dataset Bias in Medical Imaging: A Case Study on Chest X-rays
Recent work has revisited the infamous task Name that dataset and established that in non-medical datasets, there is an underlying bias and achieved high Accuracies on the dataset origin task. In this work, we revisit the same task applied to popular open-source chest X-ray datasets. Medical images are naturally more difficult to release for open-source due to their sensitive nature, which has led to certain open-source datasets being extremely popular for research purposes. By performing the same task, we wish to explore whether dataset bias also exists in these datasets. % We deliberately try to increase the difficulty of the task by dataset transformations. We apply simple transformations of the datasets to try to identify bias. Given the importance of AI applications in medical imaging, it's vital to establish whether modern methods are taking shortcuts or are focused on the relevant pathology. We implement a range of different network architectures on the datasets: NIH, CheXpert, MIMIC-CXR and PadChest. We hope this work will encourage more explainable research being performed in medical imaging and the creation of more open-source datasets in the medical domain. The corresponding code will be released upon acceptance.
☆ Breast Ultrasound Tumor Generation via Mask Generator and Text-Guided Network:A Clinically Controllable Framework with Downstream Evaluation
The development of robust deep learning models for breast ultrasound (BUS) image analysis is significantly constrained by the scarcity of expert-annotated data. To address this limitation, we propose a clinically controllable generative framework for synthesizing BUS images. This framework integrates clinical descriptions with structural masks to generate tumors, enabling fine-grained control over tumor characteristics such as morphology, echogencity, and shape. Furthermore, we design a semantic-curvature mask generator, which synthesizes structurally diverse tumor masks guided by clinical priors. During inference, synthetic tumor masks serve as input to the generative framework, producing highly personalized synthetic BUS images with tumors that reflect real-world morphological diversity. Quantitative evaluations on six public BUS datasets demonstrate the significant clinical utility of our synthetic images, showing their effectiveness in enhancing downstream breast cancer diagnosis tasks. Furthermore, visual Turing tests conducted by experienced sonographers confirm the realism of the generated images, indicating the framework's potential to support broader clinical applications.
comment: 11 pages, 6 figures
☆ Balancing the Past and Present: A Coordinated Replay Framework for Federated Class-Incremental Learning
Federated Class Incremental Learning (FCIL) aims to collaboratively process continuously increasing incoming tasks across multiple clients. Among various approaches, data replay has become a promising solution, which can alleviate forgetting by reintroducing representative samples from previous tasks. However, their performance is typically limited by class imbalance, both within the replay buffer due to limited global awareness and between replayed and newly arrived classes. To address this issue, we propose a class wise balancing data replay method for FCIL (FedCBDR), which employs a global coordination mechanism for class-level memory construction and reweights the learning objective to alleviate the aforementioned imbalances. Specifically, FedCBDR has two key components: 1) the global-perspective data replay module reconstructs global representations of prior task in a privacy-preserving manner, which then guides a class-aware and importance-sensitive sampling strategy to achieve balanced replay; 2) Subsequently, to handle class imbalance across tasks, the task aware temperature scaling module adaptively adjusts the temperature of logits at both class and instance levels based on task dynamics, which reduces the model's overconfidence in majority classes while enhancing its sensitivity to minority classes. Experimental results verified that FedCBDR achieves balanced class-wise sampling under heterogeneous data distributions and improves generalization under task imbalance between earlier and recent tasks, yielding a 2%-15% Top-1 accuracy improvement over six state-of-the-art methods.
☆ One Object, Multiple Lies: A Benchmark for Cross-task Adversarial Attack on Unified Vision-Language Models
Unified vision-language models(VLMs) have recently shown remarkable progress, enabling a single model to flexibly address diverse tasks through different instructions within a shared computational architecture. This instruction-based control mechanism creates unique security challenges, as adversarial inputs must remain effective across multiple task instructions that may be unpredictably applied to process the same malicious content. In this paper, we introduce CrossVLAD, a new benchmark dataset carefully curated from MSCOCO with GPT-4-assisted annotations for systematically evaluating cross-task adversarial attacks on unified VLMs. CrossVLAD centers on the object-change objective-consistently manipulating a target object's classification across four downstream tasks-and proposes a novel success rate metric that measures simultaneous misclassification across all tasks, providing a rigorous evaluation of adversarial transferability. To tackle this challenge, we present CRAFT (Cross-task Region-based Attack Framework with Token-alignment), an efficient region-centric attack method. Extensive experiments on Florence-2 and other popular unified VLMs demonstrate that our method outperforms existing approaches in both overall cross-task attack performance and targeted object-change success rates, highlighting its effectiveness in adversarially influencing unified VLMs across diverse tasks.
☆ Motion-Aware Adaptive Pixel Pruning for Efficient Local Motion Deblurring
Local motion blur in digital images originates from the relative motion between dynamic objects and static imaging systems during exposure. Existing deblurring methods face significant challenges in addressing this problem due to their inefficient allocation of computational resources and inadequate handling of spatially varying blur patterns. To overcome these limitations, we first propose a trainable mask predictor that identifies blurred regions in the image. During training, we employ blur masks to exclude sharp regions. For inference optimization, we implement structural reparameterization by converting $3\times 3$ convolutions to computationally efficient $1\times 1$ convolutions, enabling pixel-level pruning of sharp areas to reduce computation. Second, we develop an intra-frame motion analyzer that translates relative pixel displacements into motion trajectories, establishing adaptive guidance for region-specific blur restoration. Our method is trained end-to-end using a combination of reconstruction loss, reblur loss, and mask loss guided by annotated blur masks. Extensive experiments demonstrate superior performance over state-of-the-art methods on both local and global blur datasets while reducing FLOPs by 49\% compared to SOTA models (e.g., LMD-ViT). The source code is available at https://github.com/shangwei5/M2AENet.
comment: Accepted by ACMMM 2025
☆ Compressive Imaging Reconstruction via Tensor Decomposed Multi-Resolution Grid Encoding
Compressive imaging (CI) reconstruction, such as snapshot compressive imaging (SCI) and compressive sensing magnetic resonance imaging (MRI), aims to recover high-dimensional images from low-dimensional compressed measurements. This process critically relies on learning an accurate representation of the underlying high-dimensional image. However, existing unsupervised representations may struggle to achieve a desired balance between representation ability and efficiency. To overcome this limitation, we propose Tensor Decomposed multi-resolution Grid encoding (GridTD), an unsupervised continuous representation framework for CI reconstruction. GridTD optimizes a lightweight neural network and the input tensor decomposition model whose parameters are learned via multi-resolution hash grid encoding. It inherently enjoys the hierarchical modeling ability of multi-resolution grid encoding and the compactness of tensor decomposition, enabling effective and efficient reconstruction of high-dimensional images. Theoretical analyses for the algorithm's Lipschitz property, generalization error bound, and fixed-point convergence reveal the intrinsic superiority of GridTD as compared with existing continuous representation models. Extensive experiments across diverse CI tasks, including video SCI, spectral SCI, and compressive dynamic MRI reconstruction, consistently demonstrate the superiority of GridTD over existing methods, positioning GridTD as a versatile and state-of-the-art CI reconstruction method.
☆ D-CNN and VQ-VAE Autoencoders for Compression and Denoising of Industrial X-ray Computed Tomography Images
The ever-growing volume of data in imaging sciences stemming from the advancements in imaging technologies, necessitates efficient and reliable storage solutions for such large datasets. This study investigates the compression of industrial X-ray computed tomography (XCT) data using deep learning autoencoders and examines how these compression algorithms affect the quality of the recovered data. Two network architectures with different compression rates were used, a deep convolution neural network (D-CNN) and a vector quantized variational autoencoder (VQ-VAE). The XCT data used was from a sandstone sample with a complex internal pore network. The quality of the decoded images obtained from the two different deep learning architectures with different compression rates were quantified and compared to the original input data. In addition, to improve image decoding quality metrics, we introduced a metric sensitive to edge preservation, which is crucial for three-dimensional data analysis. We showed that different architectures and compression rates are required depending on the specific characteristics needed to be preserved for later analysis. The findings presented here can aid scientists to determine the requirements and strategies for their data storage and analysis needs.
☆ Tree-Mamba: A Tree-Aware Mamba for Underwater Monocular Depth Estimation
Underwater Monocular Depth Estimation (UMDE) is a critical task that aims to estimate high-precision depth maps from underwater degraded images caused by light absorption and scattering effects in marine environments. Recently, Mamba-based methods have achieved promising performance across various vision tasks; however, they struggle with the UMDE task because their inflexible state scanning strategies fail to model the structural features of underwater images effectively. Meanwhile, existing UMDE datasets usually contain unreliable depth labels, leading to incorrect object-depth relationships between underwater images and their corresponding depth maps. To overcome these limitations, we develop a novel tree-aware Mamba method, dubbed Tree-Mamba, for estimating accurate monocular depth maps from underwater degraded images. Specifically, we propose a tree-aware scanning strategy that adaptively constructs a minimum spanning tree based on feature similarity. The spatial topological features among the tree nodes are then flexibly aggregated through bottom-up and top-down traversals, enabling stronger multi-scale feature representation capabilities. Moreover, we construct an underwater depth estimation benchmark (called BlueDepth), which consists of 38,162 underwater image pairs with reliable depth labels. This benchmark serves as a foundational dataset for training existing deep learning-based UMDE methods to learn accurate object-depth relationships. Extensive experiments demonstrate the superiority of the proposed Tree-Mamba over several leading methods in both qualitative results and quantitative evaluations with competitive computational efficiency. Code and dataset will be available at https://wyjgr.github.io/Tree-Mamba.html.
☆ Rationale-Enhanced Decoding for Multi-modal Chain-of-Thought
Large vision-language models (LVLMs) have demonstrated remarkable capabilities by integrating pre-trained vision encoders with large language models (LLMs). Similar to single-modal LLMs, chain-of-thought (CoT) prompting has been adapted for LVLMs to enhance multi-modal reasoning by generating intermediate rationales based on visual and textual inputs. While CoT is assumed to improve grounding and accuracy in LVLMs, our experiments reveal a key challenge: existing LVLMs often ignore the contents of generated rationales in CoT reasoning. To address this, we re-formulate multi-modal CoT reasoning as a KL-constrained reward maximization focused on rationale-conditional log-likelihood. As the optimal solution, we propose rationale-enhanced decoding (RED), a novel plug-and-play inference-time decoding strategy. RED harmonizes visual and rationale information by multiplying distinct image-conditional and rationale-conditional next token distributions. Extensive experiments show that RED consistently and significantly improves reasoning over standard CoT and other decoding methods across multiple benchmarks and LVLMs. Our work offers a practical and effective approach to improve both the faithfulness and accuracy of CoT reasoning in LVLMs, paving the way for more reliable rationale-grounded multi-modal systems.
comment: 17 pages, 4 figures
☆ Action Unit Enhance Dynamic Facial Expression Recognition
Dynamic Facial Expression Recognition(DFER) is a rapidly evolving field of research that focuses on the recognition of time-series facial expressions. While previous research on DFER has concentrated on feature learning from a deep learning perspective, we put forward an AU-enhanced Dynamic Facial Expression Recognition architecture, namely AU-DFER, that incorporates AU-expression knowledge to enhance the effectiveness of deep learning modeling. In particular, the contribution of the Action Units(AUs) to different expressions is quantified, and a weight matrix is designed to incorporate a priori knowledge. Subsequently, the knowledge is integrated with the learning outcomes of a conventional deep learning network through the introduction of AU loss. The design is incorporated into the existing optimal model for dynamic expression recognition for the purpose of validation. Experiments are conducted on three recent mainstream open-source approaches to DFER on the principal datasets in this field. The results demonstrate that the proposed architecture outperforms the state-of-the-art(SOTA) methods without the need for additional arithmetic and generally produces improved results. Furthermore, we investigate the potential of AU loss function redesign to address data label imbalance issues in established dynamic expression datasets. To the best of our knowledge, this is the first attempt to integrate quantified AU-expression knowledge into various DFER models. We also devise strategies to tackle label imbalance, or minor class problems. Our findings suggest that employing a diverse strategy of loss function design can enhance the effectiveness of DFER. This underscores the criticality of addressing data imbalance challenges in mainstream datasets within this domain. The source code is available at https://github.com/Cross-Innovation-Lab/AU-DFER.
☆ Attend-and-Refine: Interactive keypoint estimation and quantitative cervical vertebrae analysis for bone age assessment
In pediatric orthodontics, accurate estimation of growth potential is essential for developing effective treatment strategies. Our research aims to predict this potential by identifying the growth peak and analyzing cervical vertebra morphology solely through lateral cephalometric radiographs. We accomplish this by comprehensively analyzing cervical vertebral maturation (CVM) features from these radiographs. This methodology provides clinicians with a reliable and efficient tool to determine the optimal timings for orthodontic interventions, ultimately enhancing patient outcomes. A crucial aspect of this approach is the meticulous annotation of keypoints on the cervical vertebrae, a task often challenged by its labor-intensive nature. To mitigate this, we introduce Attend-and-Refine Network (ARNet), a user-interactive, deep learning-based model designed to streamline the annotation process. ARNet features Interaction-guided recalibration network, which adaptively recalibrates image features in response to user feedback, coupled with a morphology-aware loss function that preserves the structural consistency of keypoints. This novel approach substantially reduces manual effort in keypoint identification, thereby enhancing the efficiency and accuracy of the process. Extensively validated across various datasets, ARNet demonstrates remarkable performance and exhibits wide-ranging applicability in medical imaging. In conclusion, our research offers an effective AI-assisted diagnostic tool for assessing growth potential in pediatric orthodontics, marking a significant advancement in the field.
comment: Accepted to Medical Image Analysis (2025)
☆ MolCLIP: A Molecular-Auxiliary CLIP Framework for Identifying Drug Mechanism of Action Based on Time-Lapsed Mitochondrial Images
Drug Mechanism of Action (MoA) mainly investigates how drug molecules interact with cells, which is crucial for drug discovery and clinical application. Recently, deep learning models have been used to recognize MoA by relying on high-content and fluorescence images of cells exposed to various drugs. However, these methods focus on spatial characteristics while overlooking the temporal dynamics of live cells. Time-lapse imaging is more suitable for observing the cell response to drugs. Additionally, drug molecules can trigger cellular dynamic variations related to specific MoA. This indicates that the drug molecule modality may complement the image counterpart. This paper proposes MolCLIP, the first visual language model to combine microscopic cell video- and molecule-modalities. MolCLIP designs a molecule-auxiliary CLIP framework to guide video features in learning the distribution of the molecular latent space. Furthermore, we integrate a metric learning strategy with MolCLIP to optimize the aggregation of video features. Experimental results on the MitoDataset demonstrate that MolCLIP achieves improvements of 51.2% and 20.5% in mAP for drug identification and MoA recognition, respectively.
☆ Bridging the gap in FER: addressing age bias in deep learning
Facial Expression Recognition (FER) systems based on deep learning have achieved impressive performance in recent years. However, these models often exhibit demographic biases, particularly with respect to age, which can compromise their fairness and reliability. In this work, we present a comprehensive study of age-related bias in deep FER models, with a particular focus on the elderly population. We first investigate whether recognition performance varies across age groups, which expressions are most affected, and whether model attention differs depending on age. Using Explainable AI (XAI) techniques, we identify systematic disparities in expression recognition and attention patterns, especially for "neutral", "sadness", and "anger" in elderly individuals. Based on these findings, we propose and evaluate three bias mitigation strategies: Multi-task Learning, Multi-modal Input, and Age-weighted Loss. Our models are trained on a large-scale dataset, AffectNet, with automatically estimated age labels and validated on balanced benchmark datasets that include underrepresented age groups. Results show consistent improvements in recognition accuracy for elderly individuals, particularly for the most error-prone expressions. Saliency heatmap analysis reveals that models trained with age-aware strategies attend to more relevant facial regions for each age group, helping to explain the observed improvements. These findings suggest that age-related bias in FER can be effectively mitigated using simple training modifications, and that even approximate demographic labels can be valuable for promoting fairness in large-scale affective computing systems.
☆ T-GVC: Trajectory-Guided Generative Video Coding at Ultra-Low Bitrates
Recent advances in video generation techniques have given rise to an emerging paradigm of generative video coding, aiming to achieve semantically accurate reconstructions in Ultra-Low Bitrate (ULB) scenarios by leveraging strong generative priors. However, most existing methods are limited by domain specificity (e.g., facial or human videos) or an excessive dependence on high-level text guidance, which often fails to capture motion details and results in unrealistic reconstructions. To address these challenges, we propose a Trajectory-Guided Generative Video Coding framework (dubbed T-GVC). T-GVC employs a semantic-aware sparse motion sampling pipeline to effectively bridge low-level motion tracking with high-level semantic understanding by extracting pixel-wise motion as sparse trajectory points based on their semantic importance, not only significantly reducing the bitrate but also preserving critical temporal semantic information. In addition, by incorporating trajectory-aligned loss constraints into diffusion processes, we introduce a training-free latent space guidance mechanism to ensure physically plausible motion patterns without sacrificing the inherent capabilities of generative models. Experimental results demonstrate that our framework outperforms both traditional codecs and state-of-the-art end-to-end video compression methods under ULB conditions. Furthermore, additional experiments confirm that our approach achieves more precise motion control than existing text-guided methods, paving the way for a novel direction of generative video coding guided by geometric motion modeling.
☆ Capture Stage Environments: A Guide to Better Matting
Capture stages are high-end sources of state-of-the-art recordings for downstream applications in movies, games, and other media. One crucial step in almost all pipelines is the matting of images to isolate the captured performances from the background. While common matting algorithms deliver remarkable performance in other applications like teleconferencing and mobile entertainment, we found that they struggle significantly with the peculiarities of capture stage content. The goal of our work is to share insights into those challenges as a curated list of those characteristics along with a constructive discussion for proactive intervention and present a guideline to practitioners for an improved workflow to mitigate unresolved challenges. To this end, we also demonstrate an efficient pipeline to adapt state-of-the-art approaches to such custom setups without the need of extensive annotations, both offline and real-time. For an objective evaluation, we propose a validation methodology based on a leading diffusion model that highlights the benefits of our approach.
☆ ViLU: Learning Vision-Language Uncertainties for Failure Prediction
Reliable Uncertainty Quantification (UQ) and failure prediction remain open challenges for Vision-Language Models (VLMs). We introduce ViLU, a new Vision-Language Uncertainty quantification framework that contextualizes uncertainty estimates by leveraging all task-relevant textual representations. ViLU constructs an uncertainty-aware multi-modal representation by integrating the visual embedding, the predicted textual embedding, and an image-conditioned textual representation via cross-attention. Unlike traditional UQ methods based on loss prediction, ViLU trains an uncertainty predictor as a binary classifier to distinguish correct from incorrect predictions using a weighted binary cross-entropy loss, making it loss-agnostic. In particular, our proposed approach is well-suited for post-hoc settings, where only vision and text embeddings are available without direct access to the model itself. Extensive experiments on diverse datasets show the significant gains of our method compared to state-of-the-art failure prediction methods. We apply our method to standard classification datasets, such as ImageNet-1k, as well as large-scale image-caption datasets like CC12M and LAION-400M. Ablation studies highlight the critical role of our architecture and training in achieving effective uncertainty quantification. Our code is publicly available and can be found here: https://github.com/ykrmm/ViLU.
☆ SpatialViz-Bench: Automatically Generated Spatial Visualization Reasoning Tasks for MLLMs
Humans can directly imagine and manipulate visual images in their minds, a capability known as spatial visualization. While multi-modal Large Language Models (MLLMs) support imagination-based reasoning, spatial visualization remains insufficiently evaluated, typically embedded within broader mathematical and logical assessments. Existing evaluations often rely on IQ tests or math competitions that may overlap with training data, compromising assessment reliability. To this end, we introduce SpatialViz-Bench, a comprehensive multi-modal benchmark for spatial visualization with 12 tasks across 4 sub-abilities, comprising 1,180 automatically generated problems. Our evaluation of 33 state-of-the-art MLLMs not only reveals wide performance variations and demonstrates the benchmark's strong discriminative power, but also uncovers counter-intuitive findings: models exhibit unexpected behaviors by showing difficulty perception that misaligns with human intuition, displaying dramatic 2D-to-3D performance cliffs, and defaulting to formula derivation despite spatial tasks requiring visualization alone. SpatialVizBench empirically demonstrates that state-of-the-art MLLMs continue to exhibit deficiencies in spatial visualization tasks, thereby addressing a significant lacuna in the field. The benchmark is publicly available.
☆ LOSC: LiDAR Open-voc Segmentation Consolidator
We study the use of image-based Vision-Language Models (VLMs) for open-vocabulary segmentation of lidar scans in driving settings. Classically, image semantics can be back-projected onto 3D point clouds. Yet, resulting point labels are noisy and sparse. We consolidate these labels to enforce both spatio-temporal consistency and robustness to image-level augmentations. We then train a 3D network based on these refined labels. This simple method, called LOSC, outperforms the SOTA of zero-shot open-vocabulary semantic and panoptic segmentation on both nuScenes and SemanticKITTI, with significant margins.
☆ HiM2SAM: Enhancing SAM2 with Hierarchical Motion Estimation and Memory Optimization towards Long-term Tracking
This paper presents enhancements to the SAM2 framework for video object tracking task, addressing challenges such as occlusions, background clutter, and target reappearance. We introduce a hierarchical motion estimation strategy, combining lightweight linear prediction with selective non-linear refinement to improve tracking accuracy without requiring additional training. In addition, we optimize the memory bank by distinguishing long-term and short-term memory frames, enabling more reliable tracking under long-term occlusions and appearance changes. Experimental results show consistent improvements across different model scales. Our method achieves state-of-the-art performance on LaSOT and LaSOText with the large model, achieving 9.6% and 7.2% relative improvements in AUC over the original SAM2, and demonstrates even larger relative gains on smaller models, highlighting the effectiveness of our trainless, low-overhead improvements for boosting long-term tracking performance. The code is available at https://github.com/LouisFinner/HiM2SAM.
☆ Stable-Hair v2: Real-World Hair Transfer via Multiple-View Diffusion Model
While diffusion-based methods have shown impressive capabilities in capturing diverse and complex hairstyles, their ability to generate consistent and high-quality multi-view outputs -- crucial for real-world applications such as digital humans and virtual avatars -- remains underexplored. In this paper, we propose Stable-Hair v2, a novel diffusion-based multi-view hair transfer framework. To the best of our knowledge, this is the first work to leverage multi-view diffusion models for robust, high-fidelity, and view-consistent hair transfer across multiple perspectives. We introduce a comprehensive multi-view training data generation pipeline comprising a diffusion-based Bald Converter, a data-augment inpainting model, and a face-finetuned multi-view diffusion model to generate high-quality triplet data, including bald images, reference hairstyles, and view-aligned source-bald pairs. Our multi-view hair transfer model integrates polar-azimuth embeddings for pose conditioning and temporal attention layers to ensure smooth transitions between views. To optimize this model, we design a novel multi-stage training strategy consisting of pose-controllable latent IdentityNet training, hair extractor training, and temporal attention training. Extensive experiments demonstrate that our method accurately transfers detailed and realistic hairstyles to source subjects while achieving seamless and consistent results across views, significantly outperforming existing methods and establishing a new benchmark in multi-view hair transfer. Code is publicly available at https://github.com/sunkymepro/StableHairV2.
comment: 14 pages
☆ HOTA: Hierarchical Overlap-Tiling Aggregation for Large-Area 3D Flood Mapping
Floods are among the most frequent natural hazards and cause significant social and economic damage. Timely, large-scale information on flood extent and depth is essential for disaster response; however, existing products often trade spatial detail for coverage or ignore flood depth altogether. To bridge this gap, this work presents HOTA: Hierarchical Overlap-Tiling Aggregation, a plug-and-play, multi-scale inference strategy. When combined with SegFormer and a dual-constraint depth estimation module, this approach forms a complete 3D flood-mapping pipeline. HOTA applies overlapping tiles of different sizes to multispectral Sentinel-2 images only during inference, enabling the SegFormer model to capture both local features and kilometre-scale inundation without changing the network weights or retraining. The subsequent depth module is based on a digital elevation model (DEM) differencing method, which refines the 2D mask and estimates flood depth by enforcing (i) zero depth along the flood boundary and (ii) near-constant flood volume with respect to the DEM. A case study on the March 2021 Kempsey (Australia) flood shows that HOTA, when coupled with SegFormer, improves IoU from 73\% (U-Net baseline) to 84\%. The resulting 3D surface achieves a mean absolute boundary error of less than 0.5 m. These results demonstrate that HOTA can produce accurate, large-area 3D flood maps suitable for rapid disaster response.
☆ NexViTAD: Few-shot Unsupervised Cross-Domain Defect Detection via Vision Foundation Models and Multi-Task Learning
This paper presents a novel few-shot cross-domain anomaly detection framework, Nexus Vision Transformer for Anomaly Detection (NexViTAD), based on vision foundation models, which effectively addresses domain-shift challenges in industrial anomaly detection through innovative shared subspace projection mechanisms and multi-task learning (MTL) module. The main innovations include: (1) a hierarchical adapter module that adaptively fuses complementary features from Hiera and DINO-v2 pre-trained models, constructing more robust feature representations; (2) a shared subspace projection strategy that enables effective cross-domain knowledge transfer through bottleneck dimension constraints and skip connection mechanisms; (3) a MTL Decoder architecture supports simultaneous processing of multiple source domains, significantly enhancing model generalization capabilities; (4) an anomaly score inference method based on Sinkhorn-K-means clustering, combined with Gaussian filtering and adaptive threshold processing for precise pixel level. Valuated on the MVTec AD dataset, NexViTAD delivers state-of-the-art performance with an AUC of 97.5%, AP of 70.4%, and PRO of 95.2% in the target domains, surpassing other recent models, marking a transformative advance in cross-domain defect detection.
☆ Diffusion-Guided Knowledge Distillation for Weakly-Supervised Low-Light Semantic Segmentation
Weakly-supervised semantic segmentation aims to assign category labels to each pixel using weak annotations, significantly reducing manual annotation costs. Although existing methods have achieved remarkable progress in well-lit scenarios, their performance significantly degrades in low-light environments due to two fundamental limitations: severe image quality degradation (e.g., low contrast, noise, and color distortion) and the inherent constraints of weak supervision. These factors collectively lead to unreliable class activation maps and semantically ambiguous pseudo-labels, ultimately compromising the model's ability to learn discriminative feature representations. To address these problems, we propose Diffusion-Guided Knowledge Distillation for Weakly-Supervised Low-light Semantic Segmentation (DGKD-WLSS), a novel framework that synergistically combines Diffusion-Guided Knowledge Distillation (DGKD) with Depth-Guided Feature Fusion (DGF2). DGKD aligns normal-light and low-light features via diffusion-based denoising and knowledge distillation, while DGF2 integrates depth maps as illumination-invariant geometric priors to enhance structural feature learning. Extensive experiments demonstrate the effectiveness of DGKD-WLSS, which achieves state-of-the-art performance in weakly supervised semantic segmentation tasks under low-light conditions. The source codes have been released at:https://github.com/ChunyanWang1/DGKD-WLSS.
☆ Beyond the Linear Separability Ceiling
Most state-of-the-art Visual-Language Models (VLMs) are seemingly limited by the linear separabilty of their visual embeddings on abstract reasoning tasks. This work investigates this "linear reasoning bottleneck" by introducing the Linear Separability Ceiling (LSC), the performance of a simple linear classifier on a VLM's visual embeddings. We find this bottleneck is widespread and stems not from poor perception, but from failures in the language model's reasoning pathways. We demonstrate this is a solvable alignment issue. The required intervention, however, is task-dependent: activating existing pathways suffices for semantic concepts, while complex relational reasoning requires adapting core model weights. Using postfix tuning as a methodological control, we find strong evidence for powerful, dormant reasoning pathways within VLMs. However, for complex relational tasks requiring deeper adaptation, explicitly improving representation quality causes the model to fail on new prompt formats despite its embeddings remaining well separated. Ultimately, this work provides a new lens for VLM analysis, showing that robust reasoning is a matter of targeted alignment, not simply improved representation learning.
☆ Single-to-mix Modality Alignment with Multimodal Large Language Model for Document Image Machine Translation ACL 2025
Document Image Machine Translation (DIMT) aims to translate text within document images, facing generalization challenges due to limited training data and the complex interplay between visual and textual information. To address these challenges, we introduce M4Doc, a novel single-to-mix modality alignment framework leveraging Multimodal Large Language Models (MLLMs). M4Doc aligns an image-only encoder with the multimodal representations of an MLLM, pre-trained on large-scale document image datasets. This alignment enables a lightweight DIMT model to learn crucial visual-textual correlations during training. During inference, M4Doc bypasses the MLLM, maintaining computational efficiency while benefiting from its multimodal knowledge. Comprehensive experiments demonstrate substantial improvements in translation quality, especially in cross-domain generalization and challenging document image scenarios.
comment: Accepted by ACL 2025 Main
☆ MAPEX: Modality-Aware Pruning of Experts for Remote Sensing Foundation Models
Remote sensing data is commonly used for tasks such as flood mapping, wildfire detection, or land-use studies. For each task, scientists carefully choose appropriate modalities or leverage data from purpose-built instruments. Recent work on remote sensing foundation models pre-trains computer vision models on large amounts of remote sensing data. These large-scale models tend to focus on specific modalities, often optical RGB or multispectral data. For many important applications, this introduces a mismatch between the application modalities and the pre-training data. Moreover, the large size of foundation models makes them expensive and difficult to fine-tune on typically small datasets for each task. We address this mismatch with MAPEX, a remote sensing foundation model based on mixture-of-modality experts. MAPEX is pre-trained on multi-modal remote sensing data with a novel modality-conditioned token routing mechanism that elicits modality-specific experts. To apply the model on a specific task, we propose a modality aware pruning technique, which only retains experts specialized for the task modalities. This yields efficient modality-specific models while simplifying fine-tuning and deployment for the modalities of interest. We experimentally validate MAPEX on diverse remote sensing datasets and show strong performance compared to fully supervised training and state-of-the-art remote sensing foundation models. Code is available at https://github.com/HSG-AIML/MAPEX.
☆ Spline Deformation Field
Trajectory modeling of dense points usually employs implicit deformation fields, represented as neural networks that map coordinates to relate canonical spatial positions to temporal offsets. However, the inductive biases inherent in neural networks can hinder spatial coherence in ill-posed scenarios. Current methods focus either on enhancing encoding strategies for deformation fields, often resulting in opaque and less intuitive models, or adopt explicit techniques like linear blend skinning, which rely on heuristic-based node initialization. Additionally, the potential of implicit representations for interpolating sparse temporal signals remains under-explored. To address these challenges, we propose a spline-based trajectory representation, where the number of knots explicitly determines the degrees of freedom. This approach enables efficient analytical derivation of velocities, preserving spatial coherence and accelerations, while mitigating temporal fluctuations. To model knot characteristics in both spatial and temporal domains, we introduce a novel low-rank time-variant spatial encoding, replacing conventional coupled spatiotemporal techniques. Our method demonstrates superior performance in temporal interpolation for fitting continuous fields with sparse inputs. Furthermore, it achieves competitive dynamic scene reconstruction quality compared to state-of-the-art methods while enhancing motion coherence without relying on linear blend skinning or as-rigid-as-possible constraints.
☆ MUVOD: A Novel Multi-view Video Object Segmentation Dataset and A Benchmark for 3D Segmentation
The application of methods based on Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3D GS) have steadily gained popularity in the field of 3D object segmentation in static scenes. These approaches demonstrate efficacy in a range of 3D scene understanding and editing tasks. Nevertheless, the 4D object segmentation of dynamic scenes remains an underexplored field due to the absence of a sufficiently extensive and accurately labelled multi-view video dataset. In this paper, we present MUVOD, a new multi-view video dataset for training and evaluating object segmentation in reconstructed real-world scenarios. The 17 selected scenes, describing various indoor or outdoor activities, are collected from different sources of datasets originating from various types of camera rigs. Each scene contains a minimum of 9 views and a maximum of 46 views. We provide 7830 RGB images (30 frames per video) with their corresponding segmentation mask in 4D motion, meaning that any object of interest in the scene could be tracked across temporal frames of a given view or across different views belonging to the same camera rig. This dataset, which contains 459 instances of 73 categories, is intended as a basic benchmark for the evaluation of multi-view video segmentation methods. We also present an evaluation metric and a baseline segmentation approach to encourage and evaluate progress in this evolving field. Additionally, we propose a new benchmark for 3D object segmentation task with a subset of annotated multi-view images selected from our MUVOD dataset. This subset contains 50 objects of different conditions in different scenarios, providing a more comprehensive analysis of state-of-the-art 3D object segmentation methods. Our proposed MUVOD dataset is available at https://volumetric-repository.labs.b-com.com/#/muvod.
☆ GGMotion: Group Graph Dynamics-Kinematics Networks for Human Motion Prediction
Human motion is a continuous physical process in 3D space, governed by complex dynamic and kinematic constraints. Existing methods typically represent the human pose as an abstract graph structure, neglecting the intrinsic physical dependencies between joints, which increases learning difficulty and makes the model prone to generating unrealistic motions. In this paper, we propose GGMotion, a group graph dynamics-kinematics network that models human topology in groups to better leverage dynamics and kinematics priors. To preserve the geometric equivariance in 3D space, we propose a novel radial field for the graph network that captures more comprehensive spatio-temporal dependencies by aggregating joint features through spatial and temporal edges. Inter-group and intra-group interaction modules are employed to capture the dependencies of joints at different scales. Combined with equivariant multilayer perceptrons (MLP), joint position features are updated in each group through parallelized dynamics-kinematics propagation to improve physical plausibility. Meanwhile, we introduce an auxiliary loss to supervise motion priors during training. Extensive experiments on three standard benchmarks, including Human3.6M, CMU-Mocap, and 3DPW, demonstrate the effectiveness and superiority of our approach, achieving a significant performance margin in short-term motion prediction. The code is available at https://github.com/inkcat520/GGMotion.git.
☆ Divergence Minimization Preference Optimization for Diffusion Model Alignment
Diffusion models have achieved remarkable success in generating realistic and versatile images from text prompts. Inspired by the recent advancements of language models, there is an increasing interest in further improving the models by aligning with human preferences. However, we investigate alignment from a divergence minimization perspective and reveal that existing preference optimization methods are typically trapped in suboptimal mean-seeking optimization. In this paper, we introduce Divergence Minimization Preference Optimization (DMPO), a novel and principled method for aligning diffusion models by minimizing reverse KL divergence, which asymptotically enjoys the same optimization direction as original RL. We provide rigorous analysis to justify the effectiveness of DMPO and conduct comprehensive experiments to validate its empirical strength across both human evaluations and automatic metrics. Our extensive results show that diffusion models fine-tuned with DMPO can consistently outperform or match existing techniques, specifically outperforming all existing diffusion alignment baselines by at least 64.6% in PickScore across all evaluation datasets, demonstrating the method's superiority in aligning generative behavior with desired outputs. Overall, DMPO unlocks a robust and elegant pathway for preference alignment, bridging principled theory with practical performance in diffusion models.
comment: 24 pages, 8 figures
☆ Semi-supervised learning and integration of multi-sequence MR-images for carotid vessel wall and plaque segmentation
The analysis of carotid arteries, particularly plaques, in multi-sequence Magnetic Resonance Imaging (MRI) data is crucial for assessing the risk of atherosclerosis and ischemic stroke. In order to evaluate metrics and radiomic features, quantifying the state of atherosclerosis, accurate segmentation is important. However, the complex morphology of plaques and the scarcity of labeled data poses significant challenges. In this work, we address these problems and propose a semi-supervised deep learning-based approach designed to effectively integrate multi-sequence MRI data for the segmentation of carotid artery vessel wall and plaque. The proposed algorithm consists of two networks: a coarse localization model identifies the region of interest guided by some prior knowledge on the position and number of carotid arteries, followed by a fine segmentation model for precise delineation of vessel walls and plaques. To effectively integrate complementary information across different MRI sequences, we investigate different fusion strategies and introduce a multi-level multi-sequence version of U-Net architecture. To address the challenges of limited labeled data and the complexity of carotid artery MRI, we propose a semi-supervised approach that enforces consistency under various input transformations. Our approach is evaluated on 52 patients with arteriosclerosis, each with five MRI sequences. Comprehensive experiments demonstrate the effectiveness of our approach and emphasize the role of fusion point selection in U-Net-based architectures. To validate the accuracy of our results, we also include an expert-based assessment of model performance. Our findings highlight the potential of fusion strategies and semi-supervised learning for improving carotid artery segmentation in data-limited MRI applications.
☆ Driving by Hybrid Navigation: An Online HD-SD Map Association Framework and Benchmark for Autonomous Vehicles
Autonomous vehicles rely on global standard-definition (SD) maps for road-level route planning and online local high-definition (HD) maps for lane-level navigation. However, recent work concentrates on construct online HD maps, often overlooking the association of global SD maps with online HD maps for hybrid navigation, making challenges in utilizing online HD maps in the real world. Observing the lack of the capability of autonomous vehicles in navigation, we introduce \textbf{O}nline \textbf{M}ap \textbf{A}ssociation, the first benchmark for the association of hybrid navigation-oriented online maps, which enhances the planning capabilities of autonomous vehicles. Based on existing datasets, the OMA contains 480k of roads and 260k of lane paths and provides the corresponding metrics to evaluate the performance of the model. Additionally, we propose a novel framework, named Map Association Transformer, as the baseline method, using path-aware attention and spatial attention mechanisms to enable the understanding of geometric and topological correspondences. The code and dataset can be accessed at https://github.com/WallelWan/OMA-MAT.
comment: 23 pages, 10 figures, 9 tables
☆ Resolving Token-Space Gradient Conflicts: Token Space Manipulation for Transformer-Based Multi-Task Learning ICCV 2025
Multi-Task Learning (MTL) enables multiple tasks to be learned within a shared network, but differences in objectives across tasks can cause negative transfer, where the learning of one task degrades another task's performance. While pre-trained transformers significantly improve MTL performance, their fixed network capacity and rigid structure limit adaptability. Previous dynamic network architectures attempt to address this but are inefficient as they directly convert shared parameters into task-specific ones. We propose Dynamic Token Modulation and Expansion (DTME-MTL), a framework applicable to any transformer-based MTL architecture. DTME-MTL enhances adaptability and reduces overfitting by identifying gradient conflicts in token space and applying adaptive solutions based on conflict type. Unlike prior methods that mitigate negative transfer by duplicating network parameters, DTME-MTL operates entirely in token space, enabling efficient adaptation without excessive parameter growth. Extensive experiments demonstrate that DTME-MTL consistently improves multi-task performance with minimal computational overhead, offering a scalable and effective solution for enhancing transformer-based MTL models.
comment: Accepted at ICCV 2025
☆ Temporal Unlearnable Examples: Preventing Personal Video Data from Unauthorized Exploitation by Object Tracking ICCV 2025
With the rise of social media, vast amounts of user-uploaded videos (e.g., YouTube) are utilized as training data for Visual Object Tracking (VOT). However, the VOT community has largely overlooked video data-privacy issues, as many private videos have been collected and used for training commercial models without authorization. To alleviate these issues, this paper presents the first investigation on preventing personal video data from unauthorized exploitation by deep trackers. Existing methods for preventing unauthorized data use primarily focus on image-based tasks (e.g., image classification), directly applying them to videos reveals several limitations, including inefficiency, limited effectiveness, and poor generalizability. To address these issues, we propose a novel generative framework for generating Temporal Unlearnable Examples (TUEs), and whose efficient computation makes it scalable for usage on large-scale video datasets. The trackers trained w/ TUEs heavily rely on unlearnable noises for temporal matching, ignoring the original data structure and thus ensuring training video data-privacy. To enhance the effectiveness of TUEs, we introduce a temporal contrastive loss, which further corrupts the learning of existing trackers when using our TUEs for training. Extensive experiments demonstrate that our approach achieves state-of-the-art performance in video data-privacy protection, with strong transferability across VOT models, datasets, and temporal matching tasks.
comment: Accepted by ICCV 2025
☆ SD-GS: Structured Deformable 3D Gaussians for Efficient Dynamic Scene Reconstruction
Current 4D Gaussian frameworks for dynamic scene reconstruction deliver impressive visual fidelity and rendering speed, however, the inherent trade-off between storage costs and the ability to characterize complex physical motions significantly limits the practical application of these methods. To tackle these problems, we propose SD-GS, a compact and efficient dynamic Gaussian splatting framework for complex dynamic scene reconstruction, featuring two key contributions. First, we introduce a deformable anchor grid, a hierarchical and memory-efficient scene representation where each anchor point derives multiple 3D Gaussians in its local spatiotemporal region and serves as the geometric backbone of the 3D scene. Second, to enhance modeling capability for complex motions, we present a deformation-aware densification strategy that adaptively grows anchors in under-reconstructed high-dynamic regions while reducing redundancy in static areas, achieving superior visual quality with fewer anchors. Experimental results demonstrate that, compared to state-of-the-art methods, SD-GS achieves an average of 60\% reduction in model size and an average of 100\% improvement in FPS, significantly enhancing computational efficiency while maintaining or even surpassing visual quality.
☆ Degradation-Agnostic Statistical Facial Feature Transformation for Blind Face Restoration in Adverse Weather Conditions
With the increasing deployment of intelligent CCTV systems in outdoor environments, there is a growing demand for face recognition systems optimized for challenging weather conditions. Adverse weather significantly degrades image quality, which in turn reduces recognition accuracy. Although recent face image restoration (FIR) models based on generative adversarial networks (GANs) and diffusion models have shown progress, their performance remains limited due to the lack of dedicated modules that explicitly address weather-induced degradations. This leads to distorted facial textures and structures. To address these limitations, we propose a novel GAN-based blind FIR framework that integrates two key components: local Statistical Facial Feature Transformation (SFFT) and Degradation-Agnostic Feature Embedding (DAFE). The local SFFT module enhances facial structure and color fidelity by aligning the local statistical distributions of low-quality (LQ) facial regions with those of high-quality (HQ) counterparts. Complementarily, the DAFE module enables robust statistical facial feature extraction under adverse weather conditions by aligning LQ and HQ encoder representations, thereby making the restoration process adaptive to severe weather-induced degradations. Experimental results demonstrate that the proposed degradation-agnostic SFFT model outperforms existing state-of-the-art FIR methods based on GAN and diffusion models, particularly in suppressing texture distortions and accurately reconstructing facial structures. Furthermore, both the SFFT and DAFE modules are empirically validated in enhancing structural fidelity and perceptual quality in face restoration under challenging weather scenarios.
☆ Objectomaly: Objectness-Aware Refinement for OoD Segmentation with Structural Consistency and Boundary Precision
Out-of-Distribution (OoD) segmentation is critical for safety-sensitive applications like autonomous driving. However, existing mask-based methods often suffer from boundary imprecision, inconsistent anomaly scores within objects, and false positives from background noise. We propose \textbf{\textit{Objectomaly}}, an objectness-aware refinement framework that incorporates object-level priors. Objectomaly consists of three stages: (1) Coarse Anomaly Scoring (CAS) using an existing OoD backbone, (2) Objectness-Aware Score Calibration (OASC) leveraging SAM-generated instance masks for object-level score normalization, and (3) Meticulous Boundary Precision (MBP) applying Laplacian filtering and Gaussian smoothing for contour refinement. Objectomaly achieves state-of-the-art performance on key OoD segmentation benchmarks, including SMIYC AnomalyTrack/ObstacleTrack and RoadAnomaly, improving both pixel-level (AuPRC up to 96.99, FPR$_{95}$ down to 0.07) and component-level (F1$-$score up to 83.44) metrics. Ablation studies and qualitative results on real-world driving videos further validate the robustness and generalizability of our method. Code will be released upon publication.
☆ Bluish Veil Detection and Lesion Classification using Custom Deep Learnable Layers with Explainable Artificial Intelligence (XAI)
Melanoma, one of the deadliest types of skin cancer, accounts for thousands of fatalities globally. The bluish, blue-whitish, or blue-white veil (BWV) is a critical feature for diagnosing melanoma, yet research into detecting BWV in dermatological images is limited. This study utilizes a non-annotated skin lesion dataset, which is converted into an annotated dataset using a proposed imaging algorithm based on color threshold techniques on lesion patches and color palettes. A Deep Convolutional Neural Network (DCNN) is designed and trained separately on three individual and combined dermoscopic datasets, using custom layers instead of standard activation function layers. The model is developed to categorize skin lesions based on the presence of BWV. The proposed DCNN demonstrates superior performance compared to conventional BWV detection models across different datasets. The model achieves a testing accuracy of 85.71% on the augmented PH2 dataset, 95.00% on the augmented ISIC archive dataset, 95.05% on the combined augmented (PH2+ISIC archive) dataset, and 90.00% on the Derm7pt dataset. An explainable artificial intelligence (XAI) algorithm is subsequently applied to interpret the DCNN's decision-making process regarding BWV detection. The proposed approach, coupled with XAI, significantly improves the detection of BWV in skin lesions, outperforming existing models and providing a robust tool for early melanoma diagnosis.
comment: Accepted version. Published in Computers in Biology and Medicine, 14 June 2024. DOI: 10.1016/j.compbiomed.2024.108758
☆ Dual Semantic-Aware Network for Noise Suppressed Ultrasound Video Segmentation
Ultrasound imaging is a prevalent diagnostic tool known for its simplicity and non-invasiveness. However, its inherent characteristics often introduce substantial noise, posing considerable challenges for automated lesion or organ segmentation in ultrasound video sequences. To address these limitations, we propose the Dual Semantic-Aware Network (DSANet), a novel framework designed to enhance noise robustness in ultrasound video segmentation by fostering mutual semantic awareness between local and global features. Specifically, we introduce an Adjacent-Frame Semantic-Aware (AFSA) module, which constructs a channel-wise similarity matrix to guide feature fusion across adjacent frames, effectively mitigating the impact of random noise without relying on pixel-level relationships. Additionally, we propose a Local-and-Global Semantic-Aware (LGSA) module that reorganizes and fuses temporal unconditional local features, which capture spatial details independently at each frame, with conditional global features that incorporate temporal context from adjacent frames. This integration facilitates multi-level semantic representation, significantly improving the model's resilience to noise interference. Extensive evaluations on four benchmark datasets demonstrate that DSANet substantially outperforms state-of-the-art methods in segmentation accuracy. Moreover, since our model avoids pixel-level feature dependencies, it achieves significantly higher inference FPS than video-based methods, and even surpasses some image-based models. Code can be found in \href{https://github.com/ZhouL2001/DSANet}{DSANet}
☆ Towards High-Resolution 3D Anomaly Detection: A Scalable Dataset and Real-Time Framework for Subtle Industrial Defects
In industrial point cloud analysis, detecting subtle anomalies demands high-resolution spatial data, yet prevailing benchmarks emphasize low-resolution inputs. To address this disparity, we propose a scalable pipeline for generating realistic and subtle 3D anomalies. Employing this pipeline, we developed MiniShift, the inaugural high-resolution 3D anomaly detection dataset, encompassing 2,577 point clouds, each with 500,000 points and anomalies occupying less than 1\% of the total. We further introduce Simple3D, an efficient framework integrating Multi-scale Neighborhood Descriptors (MSND) and Local Feature Spatial Aggregation (LFSA) to capture intricate geometric details with minimal computational overhead, achieving real-time inference exceeding 20 fps. Extensive evaluations on MiniShift and established benchmarks demonstrate that Simple3D surpasses state-of-the-art methods in both accuracy and speed, highlighting the pivotal role of high-resolution data and effective feature aggregation in advancing practical 3D anomaly detection.
comment: 14 pages, 8figures
☆ Corvid: Improving Multimodal Large Language Models Towards Chain-of-Thought Reasoning ICCV 2025
Recent advancements in multimodal large language models (MLLMs) have demonstrated exceptional performance in multimodal perception and understanding. However, leading open-source MLLMs exhibit significant limitations in complex and structured reasoning, particularly in tasks requiring deep reasoning for decision-making and problem-solving. In this work, we present Corvid, an MLLM with enhanced chain-of-thought (CoT) reasoning capabilities. Architecturally, Corvid incorporates a hybrid vision encoder for informative visual representation and a meticulously designed connector (GateMixer) to facilitate cross-modal alignment. To enhance Corvid's CoT reasoning capabilities, we introduce MCoT-Instruct-287K, a high-quality multimodal CoT instruction-following dataset, refined and standardized from diverse public reasoning sources. Leveraging this dataset, we fine-tune Corvid with a two-stage CoT-formatted training approach to progressively enhance its step-by-step reasoning abilities. Furthermore, we propose an effective inference-time scaling strategy that enables Corvid to mitigate over-reasoning and under-reasoning through self-verification. Extensive experiments demonstrate that Corvid outperforms existing o1-like MLLMs and state-of-the-art MLLMs with similar parameter scales, with notable strengths in mathematical reasoning and science problem-solving. Project page: https://mm-vl.github.io/corvid.
comment: ICCV 2025
☆ EPIC: Efficient Prompt Interaction for Text-Image Classification
In recent years, large-scale pre-trained multimodal models (LMMs) generally emerge to integrate the vision and language modalities, achieving considerable success in multimodal tasks, such as text-image classification. The growing size of LMMs, however, results in a significant computational cost for fine-tuning these models for downstream tasks. Hence, prompt-based interaction strategy is studied to align modalities more efficiently. In this context, we propose a novel efficient prompt-based multimodal interaction strategy, namely Efficient Prompt Interaction for text-image Classification (EPIC). Specifically, we utilize temporal prompts on intermediate layers, and integrate different modalities with similarity-based prompt interaction, to leverage sufficient information exchange between modalities. Utilizing this approach, our method achieves reduced computational resource consumption and fewer trainable parameters (about 1\% of the foundation model) compared to other fine-tuning strategies. Furthermore, it demonstrates superior performance on the UPMC-Food101 and SNLI-VE datasets, while achieving comparable performance on the MM-IMDB dataset.
comment: arXiv admin note: substantial text overlap with arXiv:2401.14856
☆ EscherNet++: Simultaneous Amodal Completion and Scalable View Synthesis through Masked Fine-Tuning and Enhanced Feed-Forward 3D Reconstruction
We propose EscherNet++, a masked fine-tuned diffusion model that can synthesize novel views of objects in a zero-shot manner with amodal completion ability. Existing approaches utilize multiple stages and complex pipelines to first hallucinate missing parts of the image and then perform novel view synthesis, which fail to consider cross-view dependencies and require redundant storage and computing for separate stages. Instead, we apply masked fine-tuning including input-level and feature-level masking to enable an end-to-end model with the improved ability to synthesize novel views and conduct amodal completion. In addition, we empirically integrate our model with other feed-forward image-to-mesh models without extra training and achieve competitive results with reconstruction time decreased by 95%, thanks to its ability to synthesize arbitrary query views. Our method's scalable nature further enhances fast 3D reconstruction. Despite fine-tuning on a smaller dataset and batch size, our method achieves state-of-the-art results, improving PSNR by 3.9 and Volume IoU by 0.28 on occluded tasks in 10-input settings, while also generalizing to real-world occluded reconstruction.
☆ Seg-Wild: Interactive Segmentation based on 3D Gaussian Splatting for Unconstrained Image Collections
Reconstructing and segmenting scenes from unconstrained photo collections obtained from the Internet is a novel but challenging task. Unconstrained photo collections are easier to get than well-captured photo collections. These unconstrained images suffer from inconsistent lighting and transient occlusions, which makes segmentation challenging. Previous segmentation methods cannot address transient occlusions or accurately restore the scene's lighting conditions. Therefore, we propose Seg-Wild, an interactive segmentation method based on 3D Gaussian Splatting for unconstrained image collections, suitable for in-the-wild scenes. We integrate multi-dimensional feature embeddings for each 3D Gaussian and calculate the feature similarity between the feature embeddings and the segmentation target to achieve interactive segmentation in the 3D scene. Additionally, we introduce the Spiky 3D Gaussian Cutter (SGC) to smooth abnormal 3D Gaussians. We project the 3D Gaussians onto a 2D plane and calculate the ratio of 3D Gaussians that need to be cut using the SAM mask. We also designed a benchmark to evaluate segmentation quality in in-the-wild scenes. Experimental results demonstrate that compared to previous methods, Seg-Wild achieves better segmentation results and reconstruction quality. Our code will be available at https://github.com/Sugar0725/Seg-Wild.
☆ Behave Your Motion: Habit-preserved Cross-category Animal Motion Transfer
Animal motion embodies species-specific behavioral habits, making the transfer of motion across categories a critical yet complex task for applications in animation and virtual reality. Existing motion transfer methods, primarily focused on human motion, emphasize skeletal alignment (motion retargeting) or stylistic consistency (motion style transfer), often neglecting the preservation of distinct habitual behaviors in animals. To bridge this gap, we propose a novel habit-preserved motion transfer framework for cross-category animal motion. Built upon a generative framework, our model introduces a habit-preservation module with category-specific habit encoder, allowing it to learn motion priors that capture distinctive habitual characteristics. Furthermore, we integrate a large language model (LLM) to facilitate the motion transfer to previously unobserved species. To evaluate the effectiveness of our approach, we introduce the DeformingThings4D-skl dataset, a quadruped dataset with skeletal bindings, and conduct extensive experiments and quantitative analyses, which validate the superiority of our proposed model.
☆ KeyRe-ID: Keypoint-Guided Person Re-Identification using Part-Aware Representation in Videos
We propose \textbf{KeyRe-ID}, a keypoint-guided video-based person re-identification framework consisting of global and local branches that leverage human keypoints for enhanced spatiotemporal representation learning. The global branch captures holistic identity semantics through Transformer-based temporal aggregation, while the local branch dynamically segments body regions based on keypoints to generate fine-grained, part-aware features. Extensive experiments on MARS and iLIDS-VID benchmarks demonstrate state-of-the-art performance, achieving 91.73\% mAP and 97.32\% Rank-1 accuracy on MARS, and 96.00\% Rank-1 and 100.0\% Rank-5 accuracy on iLIDS-VID. The code for this work will be publicly available on GitHub upon publication.
comment: 10 pages, 2 figures,
☆ ST-GRIT: Spatio-Temporal Graph Transformer For Internal Ice Layer Thickness Prediction ICIP
Understanding the thickness and variability of internal ice layers in radar imagery is crucial for monitoring snow accumulation, assessing ice dynamics, and reducing uncertainties in climate models. Radar sensors, capable of penetrating ice, provide detailed radargram images of these internal layers. In this work, we present ST-GRIT, a spatio-temporal graph transformer for ice layer thickness, designed to process these radargrams and capture the spatiotemporal relationships between shallow and deep ice layers. ST-GRIT leverages an inductive geometric graph learning framework to extract local spatial features as feature embeddings and employs a series of temporal and spatial attention blocks separately to model long-range dependencies effectively in both dimensions. Experimental evaluation on radargram data from the Greenland ice sheet demonstrates that ST-GRIT consistently outperforms current state-of-the-art methods and other baseline graph neural networks by achieving lower root mean-squared error. These results highlight the advantages of self-attention mechanisms on graphs over pure graph neural networks, including the ability to handle noise, avoid oversmoothing, and capture long-range dependencies. Moreover, the use of separate spatial and temporal attention blocks allows for distinct and robust learning of spatial relationships and temporal patterns, providing a more comprehensive and effective approach.
comment: Accepted for 2025 IEEE International Conference on Image Processing (ICIP)
☆ Multi-Scale Attention and Gated Shifting for Fine-Grained Event Spotting in Videos
Precise Event Spotting (PES) in sports videos requires frame-level recognition of fine-grained actions from single-camera footage. Existing PES models typically incorporate lightweight temporal modules such as Gate Shift Module (GSM) or Gate Shift Fuse (GSF) to enrich 2D CNN feature extractors with temporal context. However, these modules are limited in both temporal receptive field and spatial adaptability. We propose a Multi-Scale Attention Gate Shift Module (MSAGSM) that enhances GSM with multi-scale temporal dilations and multi-head spatial attention, enabling efficient modeling of both short- and long-term dependencies while focusing on salient regions. MSAGSM is a lightweight plug-and-play module that can be easily integrated with various 2D backbones. To further advance the field, we introduce the Table Tennis Australia (TTA) dataset-the first PES benchmark for table tennis-containing over 4800 precisely annotated events. Extensive experiments across five PES benchmarks demonstrate that MSAGSM consistently improves performance with minimal overhead, setting new state-of-the-art results.
☆ Adaptive Particle-Based Shape Modeling for Anatomical Surface Correspondence
Particle-based shape modeling (PSM) is a family of approaches that automatically quantifies shape variability across anatomical cohorts by positioning particles (pseudo landmarks) on shape surfaces in a consistent configuration. Recent advances incorporate implicit radial basis function representations as self-supervised signals to better capture the complex geometric properties of anatomical structures. However, these methods still lack self-adaptivity -- that is, the ability to automatically adjust particle configurations to local geometric features of each surface, which is essential for accurately representing complex anatomical variability. This paper introduces two mechanisms to increase surface adaptivity while maintaining consistent particle configurations: (1) a novel neighborhood correspondence loss to enable high adaptivity and (2) a geodesic correspondence algorithm that regularizes optimization to enforce geodesic neighborhood consistency. We evaluate the efficacy and scalability of our approach on challenging datasets, providing a detailed analysis of the adaptivity-correspondence trade-off and benchmarking against existing methods on surface representation accuracy and correspondence metrics.
☆ PacGDC: Label-Efficient Generalizable Depth Completion with Projection Ambiguity and Consistency ICCV 2025
Generalizable depth completion enables the acquisition of dense metric depth maps for unseen environments, offering robust perception capabilities for various downstream tasks. However, training such models typically requires large-scale datasets with metric depth labels, which are often labor-intensive to collect. This paper presents PacGDC, a label-efficient technique that enhances data diversity with minimal annotation effort for generalizable depth completion. PacGDC builds on novel insights into inherent ambiguities and consistencies in object shapes and positions during 2D-to-3D projection, allowing the synthesis of numerous pseudo geometries for the same visual scene. This process greatly broadens available geometries by manipulating scene scales of the corresponding depth maps. To leverage this property, we propose a new data synthesis pipeline that uses multiple depth foundation models as scale manipulators. These models robustly provide pseudo depth labels with varied scene scales, affecting both local objects and global layouts, while ensuring projection consistency that supports generalization. To further diversify geometries, we incorporate interpolation and relocation strategies, as well as unlabeled images, extending the data coverage beyond the individual use of foundation models. Extensive experiments show that PacGDC achieves remarkable generalizability across multiple benchmarks, excelling in diverse scene semantics/scales and depth sparsity/patterns under both zero-shot and few-shot settings. Code: https://github.com/Wang-xjtu/PacGDC.
comment: Accepted to ICCV 2025
♻ ☆ E3D-Bench: A Benchmark for End-to-End 3D Geometric Foundation Models
Spatial intelligence, encompassing 3D reconstruction, perception, and reasoning, is fundamental to applications such as robotics, aerial imaging, and extended reality. A key enabler is the real-time, accurate estimation of core 3D attributes (camera parameters, point clouds, depth maps, and 3D point tracks) from unstructured or streaming imagery. Inspired by the success of large foundation models in language and 2D vision, a new class of end-to-end 3D geometric foundation models (GFMs) has emerged, directly predicting dense 3D representations in a single feed-forward pass, eliminating the need for slow or unavailable precomputed camera parameters. Since late 2023, the field has exploded with diverse variants, but systematic evaluation is lacking. In this work, we present the first comprehensive benchmark for 3D GFMs, covering five core tasks: sparse-view depth estimation, video depth estimation, 3D reconstruction, multi-view pose estimation, novel view synthesis, and spanning both standard and challenging out-of-distribution datasets. Our standardized toolkit automates dataset handling, evaluation protocols, and metric computation to ensure fair, reproducible comparisons. We evaluate 16 state-of-the-art GFMs, revealing their strengths and limitations across tasks and domains, and derive key insights to guide future model scaling and optimization. All code, evaluation scripts, and processed data will be publicly released to accelerate research in 3D spatial intelligence.
comment: Project Page: https://e3dbench.github.io/
♻ ☆ STAR-R1: Spatial TrAnsformation Reasoning by Reinforcing Multimodal LLMs
Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities across diverse tasks, yet they lag significantly behind humans in spatial reasoning. We investigate this gap through Transformation-Driven Visual Reasoning (TVR), a challenging task requiring identification of object transformations across images under varying viewpoints. While traditional Supervised Fine-Tuning (SFT) fails to generate coherent reasoning paths in cross-view settings, sparse-reward Reinforcement Learning (RL) suffers from inefficient exploration and slow convergence. To address these limitations, we propose STAR-R1, a novel framework that integrates a single-stage RL paradigm with a fine-grained reward mechanism tailored for TVR. Specifically, STAR-R1 rewards partial correctness while penalizing excessive enumeration and passive inaction, enabling efficient exploration and precise reasoning. Comprehensive evaluations demonstrate that STAR-R1 achieves state-of-the-art performance across all 11 metrics, outperforming SFT by 23% in cross-view scenarios. Further analysis reveals STAR-R1's anthropomorphic behavior and highlights its unique ability to compare all objects for improving spatial reasoning. Our work provides critical insights in advancing the research of MLLMs and reasoning models. The codes, model weights, and data will be publicly available at https://github.com/zongzhao23/STAR-R1.
♻ ☆ Multi-modal Generative AI: Multi-modal LLMs, Diffusions and the Unification
Multi-modal generative AI (Artificial Intelligence) has attracted increasing attention from both academia and industry. Particularly, two dominant families of techniques have emerged: i) Multi-modal large language models (LLMs) demonstrate impressive ability for multi-modal understanding; and ii) Diffusion models exhibit remarkable multi-modal powers in terms of multi-modal generation. Therefore, this paper provides a comprehensive overview of multi-modal generative AI, including multi-modal LLMs, diffusions, and the unification for understanding and generation. To lay a solid foundation for unified models, we first provide a detailed review of both multi-modal LLMs and diffusion models respectively, including their probabilistic modeling procedure, multi-modal architecture design, and advanced applications to image/video LLMs as well as text-to-image/video generation. Furthermore, we explore the emerging efforts toward unified models for understanding and generation. To achieve the unification of understanding and generation, we investigate key designs including autoregressive-based and diffusion-based modeling, as well as dense and Mixture-of-Experts (MoE) architectures. We then introduce several strategies for unified models, analyzing their potential advantages and disadvantages. In addition, we summarize the common datasets widely used for multi-modal generative AI pretraining. Last but not least, we present several challenging future research directions which may contribute to the ongoing advancement of multi-modal generative AI.
comment: 20 pages, 11 figures, 2 tables
♻ ☆ Solving Inverse Problems using Diffusion with Iterative Colored Renoising
Imaging inverse problems can be solved in an unsupervised manner using pre-trained diffusion models, but doing so requires approximating the gradient of the measurement-conditional score function in the diffusion reverse process. We show that the approximations produced by existing methods are relatively poor, especially early in the reverse process, and so we propose a new approach that iteratively reestimates and "renoises" the estimate several times per diffusion step. This iterative approach, which we call Fast Iterative REnoising (FIRE), injects colored noise that is shaped to ensure that the pre-trained diffusion model always sees white noise, in accordance with how it was trained. We then embed FIRE into the DDIM reverse process and show that the resulting "DDfire" offers state-of-the-art accuracy and runtime on several linear inverse problems, as well as phase retrieval. Our implementation is at https://github.com/matt-bendel/DDfire
♻ ☆ Are Vision Transformer Representations Semantically Meaningful? A Case Study in Medical Imaging
Vision transformers (ViTs) have rapidly gained prominence in medical imaging tasks such as disease classification, segmentation, and detection due to their superior accuracy compared to conventional deep learning models. However, due to their size and complex interactions via the self-attention mechanism, they are not well understood. In particular, it is unclear whether the representations produced by such models are semantically meaningful. In this paper, using a projected gradient-based algorithm, we show that their representations are not semantically meaningful and they are inherently vulnerable to small changes. Images with imperceptible differences can have very different representations; on the other hand, images that should belong to different semantic classes can have nearly identical representations. Such vulnerability can lead to unreliable classification results; for example, unnoticeable changes cause the classification accuracy to be reduced by over 60\%. %. To the best of our knowledge, this is the first work to systematically demonstrate this fundamental lack of semantic meaningfulness in ViT representations for medical image classification, revealing a critical challenge for their deployment in safety-critical systems.
comment: 9 pages
♻ ☆ Masked Image Modeling: A Survey
In this work, we survey recent studies on masked image modeling (MIM), an approach that emerged as a powerful self-supervised learning technique in computer vision. The MIM task involves masking some information, e.g. pixels, patches, or even latent representations, and training a model, usually an autoencoder, to predicting the missing information by using the context available in the visible part of the input. We identify and formalize two categories of approaches on how to implement MIM as a pretext task, one based on reconstruction and one based on contrastive learning. Then, we construct a taxonomy and review the most prominent papers in recent years. We complement the manually constructed taxonomy with a dendrogram obtained by applying a hierarchical clustering algorithm. We further identify relevant clusters via manually inspecting the resulting dendrogram. Our review also includes datasets that are commonly used in MIM research. We aggregate the performance results of various masked image modeling methods on the most popular datasets, to facilitate the comparison of competing methods. Finally, we identify research gaps and propose several interesting directions of future work. We supplement our survey with the following public repository containing organized references: https://github.com/vladhondru25/MIM-Survey.
comment: Accepted at the International Journal of Computer Vision
♻ ☆ EyeTrAES: Fine-grained, Low-Latency Eye Tracking via Adaptive Event Slicing
Eye-tracking technology has gained significant attention in recent years due to its wide range of applications in human-computer interaction, virtual and augmented reality, and wearable health. Traditional RGB camera-based eye-tracking systems often struggle with poor temporal resolution and computational constraints, limiting their effectiveness in capturing rapid eye movements. To address these limitations, we propose EyeTrAES, a novel approach using neuromorphic event cameras for high-fidelity tracking of natural pupillary movement that shows significant kinematic variance. One of EyeTrAES's highlights is the use of a novel adaptive windowing/slicing algorithm that ensures just the right amount of descriptive asynchronous event data accumulation within an event frame, across a wide range of eye movement patterns. EyeTrAES then applies lightweight image processing functions over accumulated event frames from just a single eye to perform pupil segmentation and tracking. We show that these methods boost pupil tracking fidelity by 6+%, achieving IoU~=92%, while incurring at least 3x lower latency than competing pure event-based eye tracking alternatives [38]. We additionally demonstrate that the microscopic pupillary motion captured by EyeTrAES exhibits distinctive variations across individuals and can thus serve as a biometric fingerprint. For robust user authentication, we train a lightweight per-user Random Forest classifier using a novel feature vector of short-term pupillary kinematics, comprising a sliding window of pupil (location, velocity, acceleration) triples. Experimental studies with two different datasets demonstrate that the EyeTrAES-based authentication technique can simultaneously achieve high authentication accuracy (~=0.82) and low processing latency (~=12ms), and significantly outperform multiple state-of-the-art competitive baselines.
comment: 32 pages,15 figures,
♻ ☆ Skywork-R1V3 Technical Report
We introduce Skywork-R1V3, an advanced, open-source vision-language model (VLM) that pioneers a new approach to visual reasoning. Its key innovation lies in effectively transferring reasoning skills from text-only Large Language Models (LLMs) to visual tasks. The strong performance of Skywork-R1V3 primarily stems from our elaborate post-training RL framework, which effectively activates and enhances the model's reasoning ability, without the need for additional continue pre-training. Through this framework, we further uncover the fundamental role of the connector module in achieving robust cross-modal alignment for multimodal reasoning models. In addition, we introduce a unique indicator of reasoning capability, the entropy of critical reasoning tokens, which has proven highly effective for checkpoint selection during RL training. Skywork-R1V3 achieves state-of-the-art results on MMMU, significantly improving from 64.3% to 76.0%. This performance matches entry-level human capabilities. Remarkably, our RL-powered post-training approach enables even the 38B parameter model to rival top closed-source VLMs. The implementation successfully transfers mathematical reasoning to other subject-related reasoning tasks. We also include an analysis of curriculum learning and reinforcement finetuning strategies, along with a broader discussion on multimodal reasoning. Skywork-R1V3 represents a significant leap in multimodal reasoning, showcasing RL as a powerful engine for advancing open-source VLM capabilities.
♻ ☆ Mamba-CL: Optimizing Selective State Space Model in Null Space for Continual Learning
Continual Learning (CL) aims to equip AI models with the ability to learn a sequence of tasks over time, without forgetting previously learned knowledge. Recently, State Space Models (SSMs), particularly the Mamba model, have achieved notable success in computer vision. Building on the strengths of SSMs, this study explores leveraging the Mamba model for CL. Therefore, we introduce Mamba-CL, a framework that continuously fine-tunes the core SSMs of the large-scale Mamba foundation model by updating parameters orthogonal to the feature subspace of previous tasks. This approach theoretically guarantees the consistency objective aiming to preserves consistent output for each SSM module across both previous and current tasks, so as to overcome catastrophic forgetting issue. Specifically, we achieve this goal by deducing the overall consistency constraints on four key time-invariant parameters in the Mamba model, streamlining its recurrent state-space structure and non-linear discretization process in SSM. In practice, we apply the null-space projection to efficiently implement the orthogonality within Mamba model. Extensive experiments on four class-incremental benchmarks demonstrate the effectiveness of Mamba-CL for anti-forgetting, achieving superior performances to state-of-the-art methods. Code is available in the supplementary materials.
♻ ☆ VMem: Consistent Interactive Video Scene Generation with Surfel-Indexed View Memory
We propose a novel memory mechanism to build video generators that can explore environments interactively. Similar results have previously been achieved by out-painting 2D views of the scene while incrementally reconstructing its 3D geometry, which quickly accumulates errors, or by video generators with a short context window, which struggle to maintain scene coherence over the long term. To address these limitations, we introduce Surfel-Indexed View Memory (VMem), a mechanism that remembers past views by indexing them geometrically based on the 3D surface elements (surfels) they have observed. VMem enables the efficient retrieval of the most relevant past views when generating new ones. By focusing only on these relevant views, our method produces consistent explorations of imagined environments at a fraction of the computational cost of using all past views as context. We evaluate our approach on challenging long-term scene synthesis benchmarks and demonstrate superior performance compared to existing methods in maintaining scene coherence and camera control.
comment: Project page: https://v-mem.github.io
♻ ☆ Open-source automatic pipeline for efficient conversion of large-scale point clouds to IFC format
Building Information Modeling (BIM) is an essential component in the sustainable reconstruction and revitalization of ageing structures. However, model creation usually relies on laborious manual transformation of the unstructured point cloud data provided by laser scans or photogrammetry. This paper presents Cloud2BIM, an open-source software tool designed to automate the conversion of point clouds into BIM models compliant with the Industry Foundation Classes (IFC) standard. Cloud2BIM integrates advanced algorithms for wall and slab segmentation, opening detection, and room zoning based on real wall surfaces, resulting in a comprehensive and fully automated workflow. Unlike existing tools, it avoids computationally- and calibration-intensive techniques such as RANSAC, supports non-orthogonal geometries, and provides unprecedented processing speed-achieving results up to seven times faster than fastest competing solutions. Systematic validation using benchmark datasets confirms that Cloud2BIM is an easy-to-use, efficient, and scalable solution for generating accurate BIM models, capable of converting extensive point cloud datasets for entire buildings into IFC format with minimal user input.
comment: published version, 23 pages, 25 figures
♻ ☆ Underwater Monocular Metric Depth Estimation: Real-World Benchmarks and Synthetic Fine-Tuning with Vision Foundation Models
Monocular depth estimation has recently progressed beyond ordinal depth to provide metric depth predictions. However, its reliability in underwater environments remains limited due to light attenuation and scattering, color distortion, turbidity, and the lack of high-quality metric ground truth data. In this paper, we present a comprehensive benchmark of zero-shot and fine-tuned monocular metric depth estimation models on real-world underwater datasets with metric depth annotations, including FLSea and SQUID. We evaluated a diverse set of state-of-the-art Vision Foundation Models across a range of underwater conditions and depth ranges. Our results show that large-scale models trained on terrestrial data (real or synthetic) are effective in in-air settings, but perform poorly underwater due to significant domain shifts. To address this, we fine-tune Depth Anything V2 with a ViT-S backbone encoder on a synthetic underwater variant of the Hypersim dataset, which we simulated using a physically based underwater image formation model. Our fine-tuned model consistently improves performance across all benchmarks and outperforms baselines trained only on the clean in-air Hypersim dataset. This study presents a detailed evaluation and visualization of monocular metric depth estimation in underwater scenes, emphasizing the importance of domain adaptation and scale-aware supervision for achieving robust and generalizable metric depth predictions using foundation models in challenging environments.
♻ ☆ VOTE: Vision-Language-Action Optimization with Trajectory Ensemble Voting
Recent large-scale Vision Language Action (VLA) models have shown superior performance in robotic manipulation tasks guided by natural language. However, their generalization remains limited when applied to novel objects or unfamiliar environments that lie outside the training distribution. To address this, many existing approaches integrate additional components such as depth estimation, segmentation, or even diffusion to improve generalization, at the cost of adding significant computation overhead, resulting in low efficiency. This motivates the exploration of efficient action prediction methods, which are independent of additional high-level visual representations or diffusion techniques. In this work, we propose VOTE, an efficient and general framework for the optimization and acceleration of VLA models. In details, we propose a novel tokenizer-free fine-tuning approach for parallel accurate action prediction, which reduces computational overhead and accelerates inference speed. Additionally, we adopt an ensemble voting strategy for the action sampling, which significantly improves model performance and enhances generalization. Experimental results show that our method achieves state-of-the-art performance with 35x faster inference and 145 Hz throughput. All the details and codes will be open-sourced.
♻ ☆ Attention-Enhanced Deep Learning Ensemble for Breast Density Classification in Mammography
Breast density assessment is a crucial component of mammographic interpretation, with high breast density (BI-RADS categories C and D) representing both a significant risk factor for developing breast cancer and a technical challenge for tumor detection. This study proposes an automated deep learning system for robust binary classification of breast density (low: A/B vs. high: C/D) using the VinDr-Mammo dataset. We implemented and compared four advanced convolutional neural networks: ResNet18, ResNet50, EfficientNet-B0, and DenseNet121, each enhanced with channel attention mechanisms. To address the inherent class imbalance, we developed a novel Combined Focal Label Smoothing Loss function that integrates focal loss, label smoothing, and class-balanced weighting. Our preprocessing pipeline incorporated advanced techniques, including contrast-limited adaptive histogram equalization (CLAHE) and comprehensive data augmentation. The individual models were combined through an optimized ensemble voting approach, achieving superior performance (AUC: 0.963, F1-score: 0.952) compared to any single model. This system demonstrates significant potential to standardize density assessments in clinical practice, potentially improving screening efficiency and early cancer detection rates while reducing inter-observer variability among radiologists.
comment: 2025 IEEE Nuclear Science Symposium, Medical Imaging Conference and Room Temperature Semiconductor Detector Conference
♻ ☆ Information-driven design of imaging systems
In modern imaging systems that computationally process raw measurements before or instead of human viewing, information content matters more than visual appearance. However, developing information estimators that can handle the complexity of real-world measurements yet remain practical enough for widespread use has proven challenging. We introduce a data-driven approach for estimating mutual information between unknown objects and their noisy measurements. Our technique fits probabilistic models to measurements and their noise processes, quantifying information content without requiring ground truth data or making assumptions about object structure. We validate our approach across diverse applications-color photography, radio astronomy, lensless imaging, and microscopy-demonstrating that information estimates reliably predict system performance. Finally, we introduce Information-Driven Encoder Analysis Learning (IDEAL), which optimizes imaging systems to maximize information capture. Our work unlocks information theory as a powerful, practical tool for analyzing and designing imaging systems across a broad range of applications. A video summarizing this work can be found at: https://waller-lab.github.io/EncodingInformationWebsite/
♻ ☆ PWD: Prior-Guided and Wavelet-Enhanced Diffusion Model for Limited-Angle CT
Generative diffusion models have received increasing attention in medical imaging, particularly in limited-angle computed tomography (LACT). Standard diffusion models achieve high-quality image reconstruction but require a large number of sampling steps during inference, resulting in substantial computational overhead. Although skip-sampling strategies have been proposed to improve efficiency, they often lead to loss of fine structural details. To address this issue, we propose a prior information embedding and wavelet feature fusion fast sampling diffusion model for LACT reconstruction. The PWD enables efficient sampling while preserving reconstruction fidelity in LACT, and effectively mitigates the degradation typically introduced by skip-sampling. Specifically, during the training phase, PWD maps the distribution of LACT images to that of fully sampled target images, enabling the model to learn structural correspondences between them. During inference, the LACT image serves as an explicit prior to guide the sampling trajectory, allowing for high-quality reconstruction with significantly fewer steps. In addition, PWD performs multi-scale feature fusion in the wavelet domain, effectively enhancing the reconstruction of fine details by leveraging both low-frequency and high-frequency information. Quantitative and qualitative evaluations on clinical dental arch CBCT and periapical datasets demonstrate that PWD outperforms existing methods under the same sampling condition. Using only 50 sampling steps, PWD achieves at least 1.7 dB improvement in PSNR and 10% gain in SSIM.
♻ ☆ FunHOI: Annotation-Free 3D Hand-Object Interaction Generation via Functional Text Guidanc
Hand-object interaction(HOI) is the fundamental link between human and environment, yet its dexterous and complex pose significantly challenges for gesture control. Despite significant advances in AI and robotics, enabling machines to understand and simulate hand-object interactions, capturing the semantics of functional grasping tasks remains a considerable challenge. While previous work can generate stable and correct 3D grasps, they are still far from achieving functional grasps due to unconsidered grasp semantics. To address this challenge, we propose an innovative two-stage framework, Functional Grasp Synthesis Net (FGS-Net), for generating 3D HOI driven by functional text. This framework consists of a text-guided 3D model generator, Functional Grasp Generator (FGG), and a pose optimization strategy, Functional Grasp Refiner (FGR). FGG generates 3D models of hands and objects based on text input, while FGR fine-tunes the poses using Object Pose Approximator and energy functions to ensure the relative position between the hand and object aligns with human intent and remains physically plausible. Extensive experiments demonstrate that our approach achieves precise and high-quality HOI generation without requiring additional 3D annotation data.
♻ ☆ SVIP: Semantically Contextualized Visual Patches for Zero-Shot Learning ICCV 2025
Zero-shot learning (ZSL) aims to recognize unseen classes without labeled training examples by leveraging class-level semantic descriptors such as attributes. A fundamental challenge in ZSL is semantic misalignment, where semantic-unrelated information involved in visual features introduce ambiguity to visual-semantic interaction. Unlike existing methods that suppress semantic-unrelated information post hoc either in the feature space or the model space, we propose addressing this issue at the input stage, preventing semantic-unrelated patches from propagating through the network. To this end, we introduce Semantically contextualized VIsual Patches (SVIP) for ZSL, a transformer-based framework designed to enhance visual-semantic alignment. Specifically, we propose a self-supervised patch selection mechanism that preemptively learns to identify semantic-unrelated patches in the input space. This is trained with the supervision from aggregated attention scores across all transformer layers, which estimate each patch's semantic score. As removing semantic-unrelated patches from the input sequence may disrupt object structure, we replace them with learnable patch embeddings. With initialization from word embeddings, we can ensure they remain semantically meaningful throughout feature extraction. Extensive experiments on ZSL benchmarks demonstrate that SVIP achieves state-of-the-art performance results while providing more interpretable and semantically rich feature representations. Code is available at https://github.com/uqzhichen/SVIP.
comment: Accepted to ICCV 2025
♻ ☆ Multi-modal Representations for Fine-grained Multi-label Critical View of Safety Recognition
The Critical View of Safety (CVS) is crucial for safe laparoscopic cholecystectomy, yet assessing CVS criteria remains a complex and challenging task, even for experts. Traditional models for CVS recognition depend on vision-only models learning with costly, labor-intensive spatial annotations. This study investigates how text can be harnessed as a powerful tool for both training and inference in multi-modal surgical foundation models to automate CVS recognition. Unlike many existing multi-modal models, which are primarily adapted for multi-class classification, CVS recognition requires a multi-label framework. Zero-shot evaluation of existing multi-modal surgical models shows a significant performance gap for this task. To address this, we propose CVS-AdaptNet, a multi-label adaptation strategy that enhances fine-grained, binary classification across multiple labels by aligning image embeddings with textual descriptions of each CVS criterion using positive and negative prompts. By adapting PeskaVLP, a state-of-the-art surgical foundation model, on the Endoscapes-CVS201 dataset, CVS-AdaptNet achieves 57.6 mAP, improving over the ResNet50 image-only baseline (51.5 mAP) by 6 points. Our results show that CVS-AdaptNet's multi-label, multi-modal framework, enhanced by textual prompts, boosts CVS recognition over image-only methods. We also propose text-specific inference methods, that helps in analysing the image-text alignment. While further work is needed to match state-of-the-art spatial annotation-based methods, this approach highlights the potential of adapting generalist models to specialized surgical tasks. Code: https://github.com/CAMMA-public/CVS-AdaptNet
♻ ☆ Adaptation of Multi-modal Representation Models for Multi-task Surgical Computer Vision
Surgical AI often involves multiple tasks within a single procedure, like phase recognition or assessing the Critical View of Safety in laparoscopic cholecystectomy. Traditional models, built for one task at a time, lack flexibility, requiring a separate model for each. To address this, we introduce MML-SurgAdapt, a unified multi-task framework with Vision-Language Models (VLMs), specifically CLIP, to handle diverse surgical tasks through natural language supervision. A key challenge in multi-task learning is the presence of partial annotations when integrating different tasks. To overcome this, we employ Single Positive Multi-Label (SPML) learning, which traditionally reduces annotation burden by training models with only one positive label per instance. Our framework extends this approach to integrate data from multiple surgical tasks within a single procedure, enabling effective learning despite incomplete or noisy annotations. We demonstrate the effectiveness of our model on a combined dataset consisting of Cholec80, Endoscapes2023, and CholecT50, utilizing custom prompts. Extensive evaluation shows that MML-SurgAdapt performs comparably to task-specific benchmarks, with the added advantage of handling noisy annotations. It also outperforms the existing SPML frameworks for the task. By reducing the required labels by 23%, our approach proposes a more scalable and efficient labeling process, significantly easing the annotation burden on clinicians. To our knowledge, this is the first application of SPML to integrate data from multiple surgical tasks, presenting a novel and generalizable solution for multi-task learning in surgical computer vision. Implementation is available at: https://github.com/CAMMA-public/MML-SurgAdapt
♻ ☆ Uncertainty-Aware Gradient Stabilization for Small Object Detection
Despite advances in generic object detection, there remains a performance gap in detecting small objects compared to normal-scale objects. We reveal that conventional object localization methods suffer from gradient instability in small objects due to sharper loss curvature, leading to a convergence challenge. To address the issue, we propose Uncertainty-Aware Gradient Stabilization (UGS), a framework that reformulates object localization as a classification task to stabilize gradients. UGS quantizes continuous labels into interval non-uniform discrete representations. Under a classification-based objective, the localization branch generates bounded and confidence-driven gradients, mitigating instability. Furthermore, UGS integrates an uncertainty minimization (UM) loss that reduces prediction variance and an uncertainty-guided refinement (UR) module that identifies and refines high-uncertainty regions via perturbations. Evaluated on four benchmarks, UGS consistently improves anchor-based, anchor-free, and leading small object detectors. Especially, UGS enhances DINO-5scale by 2.6 AP on VisDrone, surpassing previous state-of-the-art results.
♻ ☆ Diffusion Model-based Data Augmentation Method for Fetal Head Ultrasound Segmentation
Medical image data is less accessible than in other domains due to privacy and regulatory constraints. In addition, labeling requires costly, time-intensive manual image annotation by clinical experts. To overcome these challenges, synthetic medical data generation offers a promising solution. Generative AI (GenAI), employing generative deep learning models, has proven effective at producing realistic synthetic images. This study proposes a novel mask-guided GenAI approach using diffusion models to generate synthetic fetal head ultrasound images paired with segmentation masks. These synthetic pairs augment real datasets for supervised fine-tuning of the Segment Anything Model (SAM). Our results show that the synthetic data captures real image features effectively, and this approach reaches state-of-the-art fetal head segmentation, especially when trained with a limited number of real image-mask pairs. In particular, the segmentation reaches Dice Scores of 94.66\% and 94.38\% using a handful of ultrasound images from the Spanish and African cohorts, respectively. Our code, models, and data are available on GitHub.
comment: Accepted at Irish Machine Vision and Image Processing Conference (IMVIP) 2025
♻ ☆ EEPNet-V2: Patch-to-Pixel Solution for Efficient Cross-Modal Registration between LiDAR Point Cloud and Camera Image
The primary requirement for cross-modal data fusion is the precise alignment of data from different sensors. However, the calibration between LiDAR point clouds and camera images is typically time-consuming and needs external calibration board or specific environmental features. Cross-modal registration effectively solves this problem by aligning the data directly without requiring external calibration. However, due to the domain gap between the point cloud and the image, existing methods rarely achieve satisfactory registration accuracy while maintaining real-time performance. To address this issue, we propose a framework that projects point clouds into several 2D representations for matching with camera images, which not only leverages the geometric characteristic of LiDAR point clouds effectively but also bridge the domain gap between the point cloud and image. Moreover, to tackle the challenges of cross modal differences and the limited overlap between LiDAR point clouds and images in the image matching task, we introduce a multi-scale feature extraction network to effectively extract features from both camera images and the projection maps of LiDAR point cloud. Additionally, we propose a patch-to-pixel matching network to provide more effective supervision and achieve high accuracy. We validate the performance of our model through experiments on the KITTI and nuScenes datasets. Experimental results demonstrate the the proposed method achieves real-time performance and extremely high registration accuracy. Specifically, on the KITTI dataset, our model achieves a registration accuracy rate of over 99\%. Our code is released at: https://github.com/ESRSchao/EEPNet-V2.
♻ ☆ HadaNorm: Diffusion Transformer Quantization through Mean-Centered Transformations
Diffusion models represent the cutting edge in image generation, but their high memory and computational demands hinder deployment on resource-constrained devices. Post-Training Quantization (PTQ) offers a promising solution by reducing the bitwidth of matrix operations. However, standard PTQ methods struggle with outliers, and achieving higher compression often requires transforming model weights and activations before quantization. In this work, we propose HadaNorm, a novel linear transformation that extends existing approaches by both normalizing channels activations and applying Hadamard transforms to effectively mitigate outliers and enable aggressive activation quantization. We demonstrate that HadaNorm consistently reduces quantization error across the various components of transformer blocks, outperforming state-of-the-art methods.
comment: 8 Pages, 6 Figures
♻ ☆ Revisiting Likelihood-Based Out-of-Distribution Detection by Modeling Representations
Out-of-distribution (OOD) detection is critical for ensuring the reliability of deep learning systems, particularly in safety-critical applications. Likelihood-based deep generative models have historically faced criticism for their unsatisfactory performance in OOD detection, often assigning higher likelihood to OOD data than in-distribution samples when applied to image data. In this work, we demonstrate that likelihood is not inherently flawed. Rather, several properties in the images space prohibit likelihood as a valid detection score. Given a sufficiently good likelihood estimator, specifically using the probability flow formulation of a diffusion model, we show that likelihood-based methods can still perform on par with state-of-the-art methods when applied in the representation space of pre-trained encoders. The code of our work can be found at $\href{https://github.com/limchaos/Likelihood-OOD.git}{\texttt{https://github.com/limchaos/Likelihood-OOD.git}}$.
comment: Scandinavian Conference on Image Analysis 2025 (oral)
♻ ☆ video-SALMONN 2: Captioning-Enhanced Audio-Visual Large Language Models
Videos contain a wealth of information, and generating detailed and accurate descriptions in natural language is a key aspect of video understanding. In this paper, we present video-SALMONN 2, an advanced audio-visual large language model (LLM) with low-rank adaptation (LoRA) designed for enhanced video (with paired audio) captioning through directed preference optimisation (DPO). We propose new metrics to evaluate the completeness and accuracy of video descriptions, which are optimised using DPO. To further improve training, we propose a novel multi-round DPO (MrDPO) approach, which involves periodically updating the DPO reference model, merging and re-initialising the LoRA module as a proxy for parameter updates after each training round (1,000 steps), and incorporating guidance from ground-truth video captions to stabilise the process. Experimental results show that MrDPO significantly enhances video-SALMONN 2's captioning accuracy, reducing the captioning error rates by 28\%. The final video-SALMONN 2 model, with just 7 billion parameters, surpasses leading models such as GPT-4o and Gemini-1.5-Pro in video captioning tasks, while maintaining highly competitive performance to the state-of-the-art on widely used video question-answering benchmarks among models of similar size. Codes are available at \href{https://github.com/bytedance/video-SALMONN-2}{https://github.com/bytedance/video-SALMONN-2}.
♻ ☆ MoSiC: Optimal-Transport Motion Trajectory for Dense Self-Supervised Learning ICCV2025
Dense self-supervised learning has shown great promise for learning pixel- and patch-level representations, but extending it to videos remains challenging due to the complexity of motion dynamics. Existing approaches struggle as they rely on static augmentations that fail under object deformations, occlusions, and camera movement, leading to inconsistent feature learning over time. We propose a motion-guided self-supervised learning framework that clusters dense point tracks to learn spatiotemporally consistent representations. By leveraging an off-the-shelf point tracker, we extract long-range motion trajectories and optimize feature clustering through a momentum-encoder-based optimal transport mechanism. To ensure temporal coherence, we propagate cluster assignments along tracked points, enforcing feature consistency across views despite viewpoint changes. Integrating motion as an implicit supervisory signal, our method learns representations that generalize across frames, improving robustness in dynamic scenes and challenging occlusion scenarios. By initializing from strong image-pretrained models and leveraging video data for training, we improve state-of-the-art by 1% to 6% on six image and video datasets and four evaluation benchmarks. The implementation is publicly available at our GitHub repository: https://github.com/SMSD75/MoSiC/tree/main
comment: Accepted to ICCV2025
♻ ☆ MedTrinity-25M: A Large-scale Multimodal Dataset with Multigranular Annotations for Medicine ICLR 2025
This paper introduces MedTrinity-25M, a comprehensive, large-scale multimodal dataset for medicine, covering over 25 million images across 10 modalities with multigranular annotations for more than 65 diseases. These multigranular annotations encompass both global information, such as modality and organ detection, and local information like ROI analysis, lesion texture, and region-wise correlations. Unlike the existing multimodal datasets, which are limited by the availability of image-text pairs, we have developed the first automated pipeline that scales up multimodal data by generating multigranular visual and textual annotations in the form of image-ROI-description triplets without the need for any paired text descriptions. Specifically, data from over 30 different sources have been collected, preprocessed, and grounded using domain-specific expert models to identify ROIs related to abnormal regions. We then build a comprehensive knowledge base and prompt multimodal large language models to perform retrieval-augmented generation with the identified ROIs as guidance, resulting in multigranular textual descriptions. Compared to existing datasets, MedTrinity-25M provides the most enriched annotations, supporting a comprehensive range of multimodal tasks such as captioning and report generation, as well as vision-centric tasks like classification and segmentation. We propose LLaVA-Tri by pretraining LLaVA on MedTrinity-25M, achieving state-of-the-art performance on VQA-RAD, SLAKE, and PathVQA, surpassing representative SOTA multimodal large language models. Furthermore, MedTrinity-25M can also be utilized to support large-scale pre-training of multimodal medical AI models, contributing to the development of future foundation models in the medical domain. We will make our dataset available.
comment: The dataset is publicly available at https://yunfeixie233.github.io/MedTrinity-25M/. Accepted to ICLR 2025
♻ ☆ Don't Get Me Wrong: How to Apply Deep Visual Interpretations to Time Series
The correct interpretation of convolutional models is a hard problem for time series data. While saliency methods promise visual validation of predictions for image and language processing, they fall short when applied to time series. These tend to be less intuitive and represent highly diverse data, such as the tool-use time series dataset. Furthermore, saliency methods often generate varied, conflicting explanations, complicating the reliability of these methods. Consequently, a rigorous objective assessment is necessary to establish trust in them. This paper investigates saliency methods on time series data to formulate recommendations for interpreting convolutional models and implements them on the tool-use time series problem. To achieve this, we first employ nine gradient-, propagation-, or perturbation-based post-hoc saliency methods across six varied and complex real-world datasets. Next, we evaluate these methods using five independent metrics to generate recommendations. Subsequently, we implement a case study focusing on tool-use time series using convolutional classification models. Our results validate our recommendations that indicate that none of the saliency methods consistently outperforms others on all metrics, while some are sometimes ahead. Our insights and step-by-step guidelines allow experts to choose suitable saliency methods for a given model and dataset.
comment: 48 pages, 12 figues, 7 tables, 6 algorithms
♻ ☆ Dance Like a Chicken: Low-Rank Stylization for Human Motion Diffusion MDM
Text-to-motion generative models span a wide range of 3D human actions but struggle with nuanced stylistic attributes such as a "Chicken" style. Due to the scarcity of style-specific data, existing approaches pull the generative prior towards a reference style, which often results in out-of-distribution low quality generations. In this work, we introduce LoRA-MDM, a lightweight framework for motion stylization that generalizes to complex actions while maintaining editability. Our key insight is that adapting the generative prior to include the style, while preserving its overall distribution, is more effective than modifying each individual motion during generation. Building on this idea, LoRA-MDM learns to adapt the prior to include the reference style using only a few samples. The style can then be used in the context of different textual prompts for generation. The low-rank adaptation shifts the motion manifold in a semantically meaningful way, enabling realistic style infusion even for actions not present in the reference samples. Moreover, preserving the distribution structure enables advanced operations such as style blending and motion editing. We compare LoRA-MDM to state-of-the-art stylized motion generation methods and demonstrate a favorable balance between text fidelity and style consistency.
comment: Project page at https://haimsaw.github.io/LoRA-MDM/
♻ ☆ Damba-ST: Domain-Adaptive Mamba for Efficient Urban Spatio-Temporal Prediction
Training urban spatio-temporal foundation models that generalize well across diverse regions and cities is critical for deploying urban services in unseen or data-scarce regions. Recent studies have typically focused on fusing cross-domain spatio-temporal data to train unified Transformer-based models. However, these models suffer from quadratic computational complexity and high memory overhead, limiting their scalability and practical deployment. Inspired by the efficiency of Mamba, a state space model with linear time complexity, we explore its potential for efficient urban spatio-temporal prediction. However, directly applying Mamba as a spatio-temporal backbone leads to negative transfer and severe performance degradation. This is primarily due to spatio-temporal heterogeneity and the recursive mechanism of Mamba's hidden state updates, which limit cross-domain generalization. To overcome these challenges, we propose Damba-ST, a novel domain-adaptive Mamba-based model for efficient urban spatio-temporal prediction. Damba-ST retains Mamba's linear complexity advantage while significantly enhancing its adaptability to heterogeneous domains. Specifically, we introduce two core innovations: (1) a domain-adaptive state space model that partitions the latent representation space into a shared subspace for learning cross-domain commonalities and independent, domain-specific subspaces for capturing intra-domain discriminative features; (2) three distinct Domain Adapters, which serve as domain-aware proxies to bridge disparate domain distributions and facilitate the alignment of cross-domain commonalities. Extensive experiments demonstrate the generalization and efficiency of Damba-ST. It achieves state-of-the-art performance on prediction tasks and demonstrates strong zero-shot generalization, enabling seamless deployment in new urban environments without extensive retraining or fine-tuning.
♻ ☆ GoalFlow: Goal-Driven Flow Matching for Multimodal Trajectories Generation in End-to-End Autonomous Driving
We propose GoalFlow, an end-to-end autonomous driving method for generating high-quality multimodal trajectories. In autonomous driving scenarios, there is rarely a single suitable trajectory. Recent methods have increasingly focused on modeling multimodal trajectory distributions. However, they suffer from trajectory selection complexity and reduced trajectory quality due to high trajectory divergence and inconsistencies between guidance and scene information. To address these issues, we introduce GoalFlow, a novel method that effectively constrains the generative process to produce high-quality, multimodal trajectories. To resolve the trajectory divergence problem inherent in diffusion-based methods, GoalFlow constrains the generated trajectories by introducing a goal point. GoalFlow establishes a novel scoring mechanism that selects the most appropriate goal point from the candidate points based on scene information. Furthermore, GoalFlow employs an efficient generative method, Flow Matching, to generate multimodal trajectories, and incorporates a refined scoring mechanism to select the optimal trajectory from the candidates. Our experimental results, validated on the Navsim\cite{Dauner2024_navsim}, demonstrate that GoalFlow achieves state-of-the-art performance, delivering robust multimodal trajectories for autonomous driving. GoalFlow achieved PDMS of 90.3, significantly surpassing other methods. Compared with other diffusion-policy-based methods, our approach requires only a single denoising step to obtain excellent performance. The code is available at https://github.com/YvanYin/GoalFlow.
♻ ☆ Mixture of Group Experts for Learning Invariant Representations
Sparsely activated Mixture-of-Experts (MoE) models effectively increase the number of parameters while maintaining consistent computational costs per token. However, vanilla MoE models often suffer from limited diversity and specialization among experts, constraining their performance and scalability, especially as the number of experts increases. In this paper, we present a novel perspective on vanilla MoE with top-$k$ routing inspired by sparse representation. This allows us to bridge established theoretical insights from sparse representation into MoE models. Building on this foundation, we propose a group sparse regularization approach for the input of top-$k$ routing, termed Mixture of Group Experts (MoGE). MoGE indirectly regularizes experts by imposing structural constraints on the routing inputs, while preserving the original MoE architecture. Furthermore, we organize the routing input into a 2D topographic map, spatially grouping neighboring elements. This structure enables MoGE to capture representations invariant to minor transformations, thereby significantly enhancing expert diversity and specialization. Comprehensive evaluations across various Transformer models for image classification and language modeling tasks demonstrate that MoGE substantially outperforms its MoE counterpart, with minimal additional memory and computation overhead. Our approach provides a simple yet effective solution to scale the number of experts and reduce redundancy among them. The source code is included in the supplementary material and will be publicly released.
♻ ☆ ReconDreamer++: Harmonizing Generative and Reconstructive Models for Driving Scene Representation
Combining reconstruction models with generative models has emerged as a promising paradigm for closed-loop simulation in autonomous driving. For example, ReconDreamer has demonstrated remarkable success in rendering large-scale maneuvers. However, a significant gap remains between the generated data and real-world sensor observations, particularly in terms of fidelity for structured elements, such as the ground surface. To address these challenges, we propose ReconDreamer++, an enhanced framework that significantly improves the overall rendering quality by mitigating the domain gap and refining the representation of the ground surface. Specifically, ReconDreamer++ introduces the Novel Trajectory Deformable Network (NTDNet), which leverages learnable spatial deformation mechanisms to bridge the domain gap between synthesized novel views and original sensor observations. Moreover, for structured elements such as the ground surface, we preserve geometric prior knowledge in 3D Gaussians, and the optimization process focuses on refining appearance attributes while preserving the underlying geometric structure. Experimental evaluations conducted on multiple datasets (Waymo, nuScenes, PandaSet, and EUVS) confirm the superior performance of ReconDreamer++. Specifically, on Waymo, ReconDreamer++ achieves performance comparable to Street Gaussians for the original trajectory while significantly outperforming ReconDreamer on novel trajectories. In particular, it achieves substantial improvements, including a 6.1% increase in NTA-IoU, a 23. 0% improvement in FID, and a remarkable 4.5% gain in the ground surface metric NTL-IoU, highlighting its effectiveness in accurately reconstructing structured elements such as the road surface.
comment: Project Page: https://recondreamer-plus.github.io/
♻ ☆ Learning to Generate Vectorized Maps at Intersections with Multiple Roadside Cameras IROS'25
Vectorized maps are indispensable for precise navigation and the safe operation of autonomous vehicles. Traditional methods for constructing these maps fall into two categories: offline techniques, which rely on expensive, labor-intensive LiDAR data collection and manual annotation, and online approaches that use onboard cameras to reduce costs but suffer from limited performance, especially at complex intersections. To bridge this gap, we introduce MRC-VMap, a cost-effective, vision-centric, end-to-end neural network designed to generate high-definition vectorized maps directly at intersections. Leveraging existing roadside surveillance cameras, MRC-VMap directly converts time-aligned, multi-directional images into vectorized map representations. This integrated solution lowers the need for additional intermediate modules--such as separate feature extraction and Bird's-Eye View (BEV) conversion steps--thus reducing both computational overhead and error propagation. Moreover, the use of multiple camera views enhances mapping completeness, mitigates occlusions, and provides robust performance under practical deployment constraints. Extensive experiments conducted on 4,000 intersections across 4 major metropolitan areas in China demonstrate that MRC-VMap not only outperforms state-of-the-art online methods but also achieves accuracy comparable to high-cost LiDAR-based approaches, thereby offering a scalable and efficient solution for modern autonomous navigation systems.
comment: Accepted by IROS'25
♻ ☆ Beyond Spatial Frequency: Pixel-wise Temporal Frequency-based Deepfake Video Detection
We introduce a deepfake video detection approach that exploits pixel-wise temporal inconsistencies, which traditional spatial frequency-based detectors often overlook. Traditional detectors represent temporal information merely by stacking spatial frequency spectra across frames, resulting in the failure to detect temporal artifacts in the pixel plane. Our approach performs a 1D Fourier transform on the time axis for each pixel, extracting features highly sensitive to temporal inconsistencies, especially in areas prone to unnatural movements. To precisely locate regions containing the temporal artifacts, we introduce an attention proposal module trained in an end-to-end manner. Additionally, our joint transformer module effectively integrates pixel-wise temporal frequency features with spatio-temporal context features, expanding the range of detectable forgery artifacts. Our framework represents a significant advancement in deepfake video detection, providing robust performance across diverse and challenging detection scenarios.
comment: accepted by iccv 2025. code is will be available at https://github.com/rama0126/PwTF-DVD
♻ ☆ GGTalker: Talking Head Systhesis with Generalizable Gaussian Priors and Identity-Specific Adaptation ICCV 2025
Creating high-quality, generalizable speech-driven 3D talking heads remains a persistent challenge. Previous methods achieve satisfactory results for fixed viewpoints and small-scale audio variations, but they struggle with large head rotations and out-of-distribution (OOD) audio. Moreover, they are constrained by the need for time-consuming, identity-specific training. We believe the core issue lies in the lack of sufficient 3D priors, which limits the extrapolation capabilities of synthesized talking heads. To address this, we propose GGTalker, which synthesizes talking heads through a combination of generalizable priors and identity-specific adaptation. We introduce a two-stage Prior-Adaptation training strategy to learn Gaussian head priors and adapt to individual characteristics. We train Audio-Expression and Expression-Visual priors to capture the universal patterns of lip movements and the general distribution of head textures. During the Customized Adaptation, individual speaking styles and texture details are precisely modeled. Additionally, we introduce a color MLP to generate fine-grained, motion-aligned textures and a Body Inpainter to blend rendered results with the background, producing indistinguishable, photorealistic video frames. Comprehensive experiments show that GGTalker achieves state-of-the-art performance in rendering quality, 3D consistency, lip-sync accuracy, and training efficiency.
comment: ICCV 2025, Project page: https://vincenthu19.github.io/GGTalker/
♻ ☆ SkipVAR: Accelerating Visual Autoregressive Modeling via Adaptive Frequency-Aware Skipping
Recent studies on Visual Autoregressive (VAR) models have highlighted that high-frequency components, or later steps, in the generation process contribute disproportionately to inference latency. However, the underlying computational redundancy involved in these steps has yet to be thoroughly investigated. In this paper, we conduct an in-depth analysis of the VAR inference process and identify two primary sources of inefficiency: step redundancy and unconditional branch redundancy. To address step redundancy, we propose an automatic step-skipping strategy that selectively omits unnecessary generation steps to improve efficiency. For unconditional branch redundancy, we observe that the information gap between the conditional and unconditional branches is minimal. Leveraging this insight, we introduce unconditional branch replacement, a technique that bypasses the unconditional branch to reduce computational cost. Notably, we observe that the effectiveness of acceleration strategies varies significantly across different samples. Motivated by this, we propose SkipVAR, a sample-adaptive framework that leverages frequency information to dynamically select the most suitable acceleration strategy for each instance. To evaluate the role of high-frequency information, we introduce high-variation benchmark datasets that test model sensitivity to fine details. Extensive experiments show SkipVAR achieves over 0.88 average SSIM with up to 1.81x overall acceleration and 2.62x speedup on the GenEval benchmark, maintaining model quality. These results confirm the effectiveness of frequency-aware, training-free adaptive acceleration for scalable autoregressive image generation. Our code is available at https://github.com/fakerone-li/SkipVAR and has been publicly released.
♻ ☆ C3T: Cross-modal Transfer Through Time for Sensor-based Human Activity Recognition
In order to unlock the potential of diverse sensors, we investigate a method to transfer knowledge between time-series modalities using a multimodal \textit{temporal} representation space for Human Activity Recognition (HAR). Specifically, we explore the setting where the modality used in testing has no labeled data during training, which we refer to as Unsupervised Modality Adaptation (UMA). We categorize existing UMA approaches as Student-Teacher or Contrastive Alignment methods. These methods typically compress continuous-time data samples into single latent vectors during alignment, inhibiting their ability to transfer temporal information through real-world temporal distortions. To address this, we introduce Cross-modal Transfer Through Time (C3T), which preserves temporal information during alignment to handle dynamic sensor data better. C3T achieves this by aligning a set of temporal latent vectors across sensing modalities. Our extensive experiments on various camera+IMU datasets demonstrate that C3T outperforms existing methods in UMA by at least 8% in accuracy and shows superior robustness to temporal distortions such as time-shift, misalignment, and dilation. Our findings suggest that C3T has significant potential for developing generalizable models for time-series sensor data, opening new avenues for various multimodal applications.
♻ ☆ Diffusion Augmented Retrieval: A Training-Free Approach to Interactive Text-to-Image Retrieval
Interactive Text-to-image retrieval (I-TIR) is an important enabler for a wide range of state-of-the-art services in domains such as e-commerce and education. However, current methods rely on finetuned Multimodal Large Language Models (MLLMs), which are costly to train and update, and exhibit poor generalizability. This latter issue is of particular concern, as: 1) finetuning narrows the pretrained distribution of MLLMs, thereby reducing generalizability; and 2) I-TIR introduces increasing query diversity and complexity. As a result, I-TIR solutions are highly likely to encounter queries and images not well represented in any training dataset. To address this, we propose leveraging Diffusion Models (DMs) for text-to-image mapping, to avoid finetuning MLLMs while preserving robust performance on complex queries. Specifically, we introduce Diffusion Augmented Retrieval (DAR), a framework that generates multiple intermediate representations via LLM-based dialogue refinements and DMs, producing a richer depiction of the user's information needs. This augmented representation facilitates more accurate identification of semantically and visually related images. Extensive experiments on four benchmarks show that for simple queries, DAR achieves results on par with finetuned I-TIR models, yet without incurring their tuning overhead. Moreover, as queries become more complex through additional conversational turns, DAR surpasses finetuned I-TIR models by up to 7.61% in Hits@10 after ten turns, illustrating its improved generalization for more intricate queries.
♻ ☆ RT-OVAD: Real-Time Open-Vocabulary Aerial Object Detection via Image-Text Collaboration
Aerial object detection plays a crucial role in numerous applications. However, most existing methods focus on detecting predefined object categories, limiting their applicability in real-world open scenarios. In this paper, we extend aerial object detection to open scenarios through image-text collaboration and propose RT-OVAD, the first real-time open-vocabulary detector for aerial scenes. Specifically, we first introduce an image-to-text alignment loss to replace the conventional category regression loss, thereby eliminating category constraints. Next, we propose a lightweight image-text collaboration strategy comprising an image-text collaboration encoder and a text-guided decoder. The encoder simultaneously enhances visual features and refines textual embeddings, while the decoder guides object queries to focus on class-relevant image features. This design further improves detection accuracy without incurring significant computational overhead. Extensive experiments demonstrate that RT-OVAD consistently outperforms existing state-of-the-art methods across open-vocabulary, zero-shot, and traditional closed-set detection tasks. For instance, on the open-vocabulary aerial detection benchmarks DIOR, DOTA-v2.0, and LAE-80C, RT-OVAD achieves 87.7 AP$_{50}$, 53.8 mAP, and 23.7 mAP, respectively, surpassing the previous state-of-the-art (LAE-DINO) by 2.2, 7.0, and 3.5 points. In addition, RT-OVAD achieves an inference speed of 34 FPS on an RTX 4090 GPU, approximately three times faster than LAE-DINO (10 FPS), meeting the real-time detection requirements of diverse applications. The code will be released at https://github.com/GT-Wei/RT-OVAD.
♻ ☆ Hybrid-View Attention Network for Clinically Significant Prostate Cancer Classification in Transrectal Ultrasound
Prostate cancer (PCa) is a leading cause of cancer-related mortality in men, and accurate identification of clinically significant PCa (csPCa) is critical for timely intervention. Transrectal ultrasound (TRUS) is widely used for prostate biopsy; however, its low contrast and anisotropic spatial resolution pose diagnostic challenges. To address these limitations, we propose a novel hybrid-view attention (HVA) network for csPCa classification in 3D TRUS that leverages complementary information from transverse and sagittal views. Our approach integrates a CNN-transformer hybrid architecture, where convolutional layers extract fine-grained local features and transformer-based HVA models global dependencies. Specifically, the HVA comprises intra-view attention to refine features within a single view and cross-view attention to incorporate complementary information across views. Furthermore, a hybrid-view adaptive fusion module dynamically aggregates features along both channel and spatial dimensions, enhancing the overall representation. Experiments are conducted on an in-house dataset containing 590 subjects who underwent prostate biopsy. Comparative and ablation results prove the efficacy of our method. The code is available at https://github.com/mock1ngbrd/HVAN.
♻ ☆ Boundary Learning by Using Weighted Propagation in Convolution Network
In material science, image segmentation is of great significance for quantitative analysis of microstructures. Here, we propose a novel Weighted Propagation Convolution Neural Network based on U-Net (WPU-Net) to detect boundary in poly-crystalline microscopic images. We introduce spatial consistency into network to eliminate the defects in raw microscopic image. And we customize adaptive boundary weight for each pixel in each grain, so that it leads the network to preserve grain's geometric and topological characteristics. Moreover, we provide our dataset with the goal of advancing the development of image processing in materials science. Experiments demonstrate that the proposed method achieves promising performance in both of objective and subjective assessment. In boundary detection task, it reduces the error rate by 7\%, which outperforms state-of-the-art methods by a large margin.
comment: technical report
♻ ☆ Concept Unlearning by Modeling Key Steps of Diffusion Process
Text-to-image diffusion models (T2I DMs), represented by Stable Diffusion, which generate highly realistic images based on textual input, have been widely used. However, their misuse poses serious security risks. While existing concept unlearning methods aim to mitigate these risks, they struggle to balance unlearning effectiveness with generative retainability.To overcome this limitation, we innovatively propose the Key Step Concept Unlearning (KSCU) method, which ingeniously capitalizes on the unique stepwise sampling characteristic inherent in diffusion models during the image generation process. Unlike conventional approaches that treat all denoising steps equally, KSCU strategically focuses on pivotal steps with the most influence over the final outcome by dividing key steps for different concept unlearning tasks and fine-tuning the model only at those steps. This targeted approach reduces the number of parameter updates needed for effective unlearning, while maximizing the retention of the model's generative capabilities.Through extensive benchmark experiments, we demonstrate that KSCU effectively prevents T2I DMs from generating undesirable images while better retaining the model's generative capabilities. Our code will be released.
♻ ☆ DLaVA: Document Language and Vision Assistant for Answer Localization with Enhanced Interpretability and Trustworthiness
Document Visual Question Answering (VQA) demands robust integration of text detection, recognition, and spatial reasoning to interpret complex document layouts. In this work, we introduce DLaVA, a novel, training-free pipeline that leverages Multimodal Large Language Models (MLLMs) for zero-shot answer localization in order to improve trustworthiness, interpretability, and explainability. By leveraging an innovative OCR-free approach that organizes text regions with unique bounding box IDs, the proposed method preserves spatial contexts without relying on iterative OCR or chain-of-thought reasoning, thus substantially reducing the computational complexity. We further enhance the evaluation protocol by integrating Intersection over Union (IoU) metrics alongside Average Normalized Levenshtein Similarity (ANLS), thereby ensuring that not only textual accuracy is considered, but spatial accuracy is taken into account, ultimately reducing the risks of AI hallucinations and improving trustworthiness. Experiments on benchmark datasets demonstrate competitive performance compared to state-of-the-art techniques, with significantly lower computational complexity and enhanced accuracies and reliability for high-stakes applications. The code and datasets utilized in this study for DLaVA are accessible at: https://github.com/ahmad-shirazi/AnnotMLLM.
♻ ☆ Online Continual Learning via Spiking Neural Networks with Sleep Enhanced Latent Replay
Edge computing scenarios necessitate the development of hardware-efficient online continual learning algorithms to be adaptive to dynamic environment. However, existing algorithms always suffer from high memory overhead and bias towards recently trained tasks. To tackle these issues, this paper proposes a novel online continual learning approach termed as SESLR, which incorporates a sleep enhanced latent replay scheme with spiking neural networks (SNNs). SESLR leverages SNNs' binary spike characteristics to store replay features in single bits, significantly reducing memory overhead. Furthermore, inspired by biological sleep-wake cycles, SESLR introduces a noise-enhanced sleep phase where the model exclusively trains on replay samples with controlled noise injection, effectively mitigating classification bias towards new classes. Extensive experiments on both conventional (MNIST, CIFAR10) and neuromorphic (NMNIST, CIFAR10-DVS) datasets demonstrate SESLR's effectiveness. On Split CIFAR10, SESLR achieves nearly 30% improvement in average accuracy with only one-third of the memory consumption compared to baseline methods. On Split CIFAR10-DVS, it improves accuracy by approximately 10% while reducing memory overhead by a factor of 32. These results validate SESLR as a promising solution for online continual learning in resource-constrained edge computing scenarios.
comment: 9 pages, 4figures
♻ ☆ Localized Concept Erasure for Text-to-Image Diffusion Models Using Training-Free Gated Low-Rank Adaptation CVPR 2025
Fine-tuning based concept erasing has demonstrated promising results in preventing generation of harmful contents from text-to-image diffusion models by removing target concepts while preserving remaining concepts. To maintain the generation capability of diffusion models after concept erasure, it is necessary to remove only the image region containing the target concept when it locally appears in an image, leaving other regions intact. However, prior arts often compromise fidelity of the other image regions in order to erase the localized target concept appearing in a specific area, thereby reducing the overall performance of image generation. To address these limitations, we first introduce a framework called localized concept erasure, which allows for the deletion of only the specific area containing the target concept in the image while preserving the other regions. As a solution for the localized concept erasure, we propose a training-free approach, dubbed Gated Low-rank adaptation for Concept Erasure (GLoCE), that injects a lightweight module into the diffusion model. GLoCE consists of low-rank matrices and a simple gate, determined only by several generation steps for concepts without training. By directly applying GLoCE to image embeddings and designing the gate to activate only for target concepts, GLoCE can selectively remove only the region of the target concepts, even when target and remaining concepts coexist within an image. Extensive experiments demonstrated GLoCE not only improves the image fidelity to text prompts after erasing the localized target concepts, but also outperforms prior arts in efficacy, specificity, and robustness by large margin and can be extended to mass concept erasure.
comment: Accepted to CVPR 2025
♻ ☆ MCFormer: A Multi-Cost-Volume Network and Comprehensive Benchmark for Particle Image Velocimetry
Particle Image Velocimetry (PIV) is fundamental to fluid dynamics, yet deep learning applications face significant hurdles. A critical gap exists: the lack of comprehensive evaluation of how diverse optical flow models perform specifically on PIV data, largely due to limitations in available datasets and the absence of a standardized benchmark. This prevents fair comparison and hinders progress. To address this, our primary contribution is a novel, large-scale synthetic PIV benchmark dataset generated from diverse CFD simulations (JHTDB and Blasius). It features unprecedented variety in particle densities, flow velocities, and continuous motion, enabling, for the first time, a standardized and rigorous evaluation of various optical flow and PIV algorithms. Complementing this, we propose Multi Cost Volume PIV (MCFormer), a new deep network architecture leveraging multi-frame temporal information and multiple cost volumes, specifically designed for PIV's sparse nature. Our comprehensive benchmark evaluation, the first of its kind, reveals significant performance variations among adapted optical flow models and demonstrates that MCFormer significantly outperforms existing methods, achieving the lowest overall normalized endpoint error (NEPE). This work provides both a foundational benchmark resource essential for future PIV research and a state-of-the-art method tailored for PIV challenges. We make our benchmark dataset and code publicly available to foster future research in this area.
comment: 20 pages, 13 figures, 5 tables. Comprehensive benchmark evaluation of optical flow models for PIV. Introduces MCFormer architecture with multi-frame temporal processing and multiple cost volumes. Includes large-scale synthetic PIV dataset based on JHTDB and Blasius CFD simulations. Code and dataset will be made publicly available
♻ ☆ Hallucinating 360°: Panoramic Street-View Generation via Local Scenes Diffusion and Probabilistic Prompting
Panoramic perception holds significant potential for autonomous driving, enabling vehicles to acquire a comprehensive 360{\deg} surround view in a single shot. However, autonomous driving is a data-driven task. Complete panoramic data acquisition requires complex sampling systems and annotation pipelines, which are time-consuming and labor-intensive. Although existing street view generation models have demonstrated strong data regeneration capabilities, they can only learn from the fixed data distribution of existing datasets and cannot achieve high-quality, controllable panoramic generation. In this paper, we propose the first panoramic generation method Percep360 for autonomous driving. Percep360 enables coherent generation of panoramic data with control signals based on the stitched panoramic data. Percep360 focuses on two key aspects: coherence and controllability. Specifically, to overcome the inherent information loss caused by the pinhole sampling process, we propose the Local Scenes Diffusion Method (LSDM). LSDM reformulates the panorama generation as a spatially continuous diffusion process, bridging the gaps between different data distributions. Additionally, to achieve the controllable generation of panoramic images, we propose a Probabilistic Prompting Method (PPM). PPM dynamically selects the most relevant control cues, enabling controllable panoramic image generation. We evaluate the effectiveness of the generated images from three perspectives: image quality assessment (i.e., no-reference and with reference), controllability, and their utility in real-world Bird's Eye View (BEV) segmentation. Notably, the generated data consistently outperforms the original stitched images in no-reference quality metrics and enhances downstream perception models. The source code will be publicly available at https://github.com/Bryant-Teng/Percep360.
comment: The source code will be publicly available at https://github.com/Bryant-Teng/Percep360
Information Retrieval 20
☆ Measuring Hypothesis Testing Errors in the Evaluation of Retrieval Systems
The evaluation of Information Retrieval (IR) systems typically uses query-document pairs with corresponding human-labelled relevance assessments (qrels). These qrels are used to determine if one system is better than another based on average retrieval performance. Acquiring large volumes of human relevance assessments is expensive. Therefore, more efficient relevance assessment approaches have been proposed, necessitating comparisons between qrels to ascertain their efficacy. Discriminative power, i.e. the ability to correctly identify significant differences between systems, is important for drawing accurate conclusions on the robustness of qrels. Previous work has measured the proportion of pairs of systems that are identified as significantly different and has quantified Type I statistical errors. Type I errors lead to incorrect conclusions due to false positive significance tests. We argue that also identifying Type II errors (false negatives) is important as they lead science in the wrong direction. We quantify Type II errors and propose that balanced classification metrics, such as balanced accuracy, can be used to portray the discriminative power of qrels. We perform experiments using qrels generated using alternative relevance assessment methods to investigate measuring hypothesis testing errors in IR evaluation. We find that additional insights into the discriminative power of qrels can be gained by quantifying Type II errors, and that balanced classification metrics can be used to give an overall summary of discriminative power in one, easily comparable, number.
☆ Plausible Counterfactual Explanations of Recommendations
Explanations play a variety of roles in various recommender systems, from a legally mandated afterthought, through an integral element of user experience, to a key to persuasiveness. A natural and useful form of an explanation is the Counterfactual Explanation (CE). We present a method for generating highly plausible CEs in recommender systems and evaluate it both numerically and with a user study.
comment: 8 pages, 3 figures, 6 tables
☆ DTECT: Dynamic Topic Explorer & Context Tracker
The explosive growth of textual data over time presents a significant challenge in uncovering evolving themes and trends. Existing dynamic topic modeling techniques, while powerful, often exist in fragmented pipelines that lack robust support for interpretation and user-friendly exploration. We introduce DTECT (Dynamic Topic Explorer & Context Tracker), an end-to-end system that bridges the gap between raw textual data and meaningful temporal insights. DTECT provides a unified workflow that supports data preprocessing, multiple model architectures, and dedicated evaluation metrics to analyze the topic quality of temporal topic models. It significantly enhances interpretability by introducing LLM-driven automatic topic labeling, trend analysis via temporally salient words, interactive visualizations with document-level summarization, and a natural language chat interface for intuitive data querying. By integrating these features into a single, cohesive platform, DTECT empowers users to more effectively track and understand thematic dynamics. DTECT is open-source and available at https://github.com/AdhyaSuman/DTECT.
comment: Code: https://github.com/AdhyaSuman/DTECT | Demo: https://huggingface.co/spaces/AdhyaSuman/DTECT | Video: https://youtu.be/B8nNfxFoJAU
☆ Document Similarity Enhanced IPS Estimation for Unbiased Learning to Rank
Learning to Rank (LTR) models learn from historical user interactions, such as user clicks. However, there is an inherent bias in the clicks of users due to position bias, i.e., users are more likely to click highly-ranked documents than low-ranked documents. To address this bias when training LTR models, many approaches from the literature re-weight the users' click data using Inverse Propensity Scoring (IPS). IPS re-weights the user's clicks proportionately to the position in the historical ranking that a document was placed when it was clicked since low-ranked documents are less likely to be seen by a user. In this paper, we argue that low-ranked documents that are similar to highly-ranked relevant documents are also likely to be relevant. Moreover, accounting for the similarity of low-ranked documents to highly ranked relevant documents when calculating IPS can more effectively mitigate the effects of position bias. Therefore, we propose an extension to IPS, called IPSsim, that takes into consideration the similarity of documents when estimating IPS. We evaluate our IPSsim estimator using two large publicly available LTR datasets under a number of simulated user click settings, and with different numbers of training clicks. Our experiments show that our IPSsim estimator is more effective than the existing IPS estimators for learning an unbiased LTR model, particularly in top-n settings when n >= 30. For example, when n = 50, our IPSsim estimator achieves a statistically significant ~3% improvement (p < 0.05) in terms of NDCG compared to the Doubly Robust estimator from the literature.
☆ Rethinking the Privacy of Text Embeddings: A Reproducibility Study of "Text Embeddings Reveal (Almost) As Much As Text" RecSys 2025
Text embeddings are fundamental to many natural language processing (NLP) tasks, extensively applied in domains such as recommendation systems and information retrieval (IR). Traditionally, transmitting embeddings instead of raw text has been seen as privacy-preserving. However, recent methods such as Vec2Text challenge this assumption by demonstrating that controlled decoding can successfully reconstruct original texts from black-box embeddings. The unexpectedly strong results reported by Vec2Text motivated us to conduct further verification, particularly considering the typically non-intuitive and opaque structure of high-dimensional embedding spaces. In this work, we reproduce the Vec2Text framework and evaluate it from two perspectives: (1) validating the original claims, and (2) extending the study through targeted experiments. First, we successfully replicate the original key results in both in-domain and out-of-domain settings, with only minor discrepancies arising due to missing artifacts, such as model checkpoints and dataset splits. Furthermore, we extend the study by conducting a parameter sensitivity analysis, evaluating the feasibility of reconstructing sensitive inputs (e.g., passwords), and exploring embedding quantization as a lightweight privacy defense. Our results show that Vec2Text is effective under ideal conditions, capable of reconstructing even password-like sequences that lack clear semantics. However, we identify key limitations, including its sensitivity to input sequence length. We also find that Gaussian noise and quantization techniques can mitigate the privacy risks posed by Vec2Text, with quantization offering a simpler and more widely applicable solution. Our findings emphasize the need for caution in using text embeddings and highlight the importance of further research into robust defense mechanisms for NLP systems.
comment: This paper has been accepted for oral presentation in the reproducibility track at RecSys 2025
☆ The Cross-Lingual Cost: Retrieval Biases in RAG over Arabic-English Corpora
Cross-lingual retrieval-augmented generation (RAG) is a critical capability for retrieving and generating answers across languages. Prior work in this context has mostly focused on generation and relied on benchmarks derived from open-domain sources, most notably Wikipedia. In such settings, retrieval challenges often remain hidden due to language imbalances, overlap with pretraining data, and memorized content. To address this gap, we study Arabic-English RAG in a domain-specific setting using benchmarks derived from real-world corporate datasets. Our benchmarks include all combinations of languages for the user query and the supporting document, drawn independently and uniformly at random. This enables a systematic study of multilingual retrieval behavior. Our findings reveal that retrieval is a critical bottleneck in cross-lingual domain-specific scenarios, with significant performance drops occurring when the user query and supporting document languages differ. A key insight is that these failures stem primarily from the retriever's difficulty in ranking documents across languages. Finally, we propose a simple retrieval strategy that addresses this source of failure by enforcing equal retrieval from both languages, resulting in substantial improvements in cross-lingual and overall performance. These results highlight meaningful opportunities for improving multilingual retrieval, particularly in practical, real-world RAG applications.
☆ NLGCL: Naturally Existing Neighbor Layers Graph Contrastive Learning for Recommendation RecSys 2025
Graph Neural Networks (GNNs) are widely used in collaborative filtering to capture high-order user-item relationships. To address the data sparsity problem in recommendation systems, Graph Contrastive Learning (GCL) has emerged as a promising paradigm that maximizes mutual information between contrastive views. However, existing GCL methods rely on augmentation techniques that introduce semantically irrelevant noise and incur significant computational and storage costs, limiting effectiveness and efficiency. To overcome these challenges, we propose NLGCL, a novel contrastive learning framework that leverages naturally contrastive views between neighbor layers within GNNs. By treating each node and its neighbors in the next layer as positive pairs, and other nodes as negatives, NLGCL avoids augmentation-based noise while preserving semantic relevance. This paradigm eliminates costly view construction and storage, making it computationally efficient and practical for real-world scenarios. Extensive experiments on four public datasets demonstrate that NLGCL outperforms state-of-the-art baselines in effectiveness and efficiency.
comment: Accepted by RecSys 2025 as Spotlight Oral
☆ When Graph Contrastive Learning Backfires: Spectral Vulnerability and Defense in Recommendation
Graph Contrastive Learning (GCL) has demonstrated substantial promise in enhancing the robustness and generalization of recommender systems, particularly by enabling models to leverage large-scale unlabeled data for improved representation learning. However, in this paper, we reveal an unexpected vulnerability: the integration of GCL inadvertently increases the susceptibility of a recommender to targeted promotion attacks. Through both theoretical investigation and empirical validation, we identify the root cause as the spectral smoothing effect induced by contrastive optimization, which disperses item embeddings across the representation space and unintentionally enhances the exposure of target items. Building on this insight, we introduce CLeaR, a bi-level optimization attack method that deliberately amplifies spectral smoothness, enabling a systematic investigation of the susceptibility of GCL-based recommendation models to targeted promotion attacks. Our findings highlight the urgent need for robust countermeasures; in response, we further propose SIM, a spectral irregularity mitigation framework designed to accurately detect and suppress targeted items without compromising model performance. Extensive experiments on multiple benchmark datasets demonstrate that, compared to existing targeted promotion attacks, GCL-based recommendation models exhibit greater susceptibility when evaluated with CLeaR, while SIM effectively mitigates these vulnerabilities.
comment: 24 pages, 6 figures
Overview of the TREC 2021 deep learning track
This is the third year of the TREC Deep Learning track. As in previous years, we leverage the MS MARCO datasets that made hundreds of thousands of human annotated training labels available for both passage and document ranking tasks. In addition, this year we refreshed both the document and the passage collections which also led to a nearly four times increase in the document collection size and nearly $16$ times increase in the size of the passage collection. Deep neural ranking models that employ large scale pretraininig continued to outperform traditional retrieval methods this year. We also found that single stage retrieval can achieve good performance on both tasks although they still do not perform at par with multistage retrieval pipelines. Finally, the increase in the collection size and the general data refresh raised some questions about completeness of NIST judgments and the quality of the training labels that were mapped to the new collections from the old ones which we discuss in this report.
☆ GRASP: Generic Reasoning And SPARQL Generation across Knowledge Graphs
We propose a new approach for generating SPARQL queries on RDF knowledge graphs from natural language questions or keyword queries, using a large language model. Our approach does not require fine-tuning. Instead, it uses the language model to explore the knowledge graph by strategically executing SPARQL queries and searching for relevant IRIs and literals. We evaluate our approach on a variety of benchmarks (for knowledge graphs of different kinds and sizes) and language models (of different scales and types, commercial as well as open-source) and compare it with existing approaches. On Wikidata we reach state-of-the-art results on multiple benchmarks, despite the zero-shot setting. On Freebase we come close to the best few-shot methods. On other, less commonly evaluated knowledge graphs and benchmarks our approach also performs well overall. We conduct several additional studies, like comparing different ways of searching the graphs, incorporating a feedback mechanism, or making use of few-shot examples.
♻ ☆ The Future is Agentic: Definitions, Perspectives, and Open Challenges of Multi-Agent Recommender Systems
Large language models (LLMs) are rapidly evolving from passive engines of text generation into agentic entities that can plan, remember, invoke external tools, and co-operate with one another. This perspective paper investigates how such LLM agents (and societies thereof) can transform the design space of recommender systems. We introduce a unified formalism that (i) models an individual agent as a tuple comprising its language core, tool set, and hierarchical memory, and (ii) captures a multi-agent recommender as a triple of agents, shared environment, and communication protocol. Within this framework, we present four end-to-end use cases-interactive party planning, synthetic user-simulation for offline evaluation, multi-modal furniture recommendation, and brand-aligned explanation generation-each illustrating a distinct capability unlocked by agentic orchestration. We then surface five cross-cutting challenge families: protocol complexity, scalability, hallucination and error propagation, emergent misalignment (including covert collusion), and brand compliance. For each, we formalize the problem, review nascent mitigation strategies, and outline open research questions. The result is both a blueprint and an agenda: a blueprint that shows how memory-augmented, tool-using LLM agents can be composed into robust recommendation pipelines, and an agenda inviting the RecSys community to develop benchmarks, theoretical guarantees, and governance tools that keep pace with this new degree of autonomy. By unifying agentic abstractions with recommender objectives, the paper lays the groundwork for the next generation of personalized, trustworthy, and context-rich recommendation services.
♻ ☆ Adaptive Graph Integration for Cross-Domain Recommendation via Heterogeneous Graph Coordinators SIGIR 2025
In the digital era, users typically interact with diverse items across multiple domains (e.g., e-commerce, streaming platforms, and social networks), generating intricate heterogeneous interaction graphs. Leveraging multi-domain data can improve recommendation systems by enriching user insights and mitigating data sparsity in individual domains. However, integrating such multi-domain knowledge for cross-domain recommendation remains challenging due to inherent disparities in user behavior and item characteristics and the risk of negative transfer, where irrelevant or conflicting information from the source domains adversely impacts the target domain's performance. To tackle these challenges, we propose HAGO, a novel framework with \textbf{H}eterogeneous \textbf{A}daptive \textbf{G}raph co\textbf{O}rdinators, which dynamically integrates multi-domain graphs into a cohesive structure. HAGO adaptively adjusts the connections between coordinators and multi-domain graph nodes to enhance beneficial inter-domain interactions while alleviating negative transfer. Furthermore, we introduce a universal multi-domain graph pre-training strategy alongside HAGO to collaboratively learn high-quality node representations across domains. Being compatible with various graph-based models and pre-training techniques, HAGO demonstrates broad applicability and effectiveness. Extensive experiments show that our framework outperforms state-of-the-art methods in cross-domain recommendation scenarios, underscoring its potential for real-world applications. The source code is available at https://github.com/zhy99426/HAGO.
comment: Accept by SIGIR 2025
♻ ☆ Contextual Bandits in Payment Processing: Non-uniform Exploration and Supervised Learning KDD '25
Uniform random exploration in decision-making systems supports off-policy learning via supervision but incurs high regret, making it impractical for many applications. Conversely, non-uniform exploration offers better immediate performance but lacks support for off-policy learning. Recent research suggests that regression oracles can bridge this gap by combining non-uniform exploration with supervised learning. In this paper, we analyze these approaches within a real-world industrial context at Adyen, a large global payments processor characterized by batch logged delayed feedback, short-term memory, and dynamic action spaces under the Empirical Risk Minimization (ERM) framework. Our analysis reveals that while regression oracles significantly improve performance, they introduce challenges due to rigid algorithmic assumptions. Specifically, we observe that as a policy improves, subsequent generations may perform worse due to shifts in the reward distribution and increased class imbalance in the training data. This degradation occurs de spite improvements in other aspects of the training data, leading to decreased performance in successive policy iterations. We further explore the long-term impact of regression oracles, identifying a potential "oscillation effect." This effect arises when regression oracles influence probability estimates and the realizability of subsequent policy models, leading to fluctuations in performance across iterations. Our findings highlight the need for more adaptable algorithms that can leverage the benefits of regression oracles without introducing instability in policy performance over time.
comment: 7 pages, 10 figures, submitted to KDD '25
♻ ☆ Toward Holistic Evaluation of Recommender Systems Powered by Generative Models
Recommender systems powered by generative models (Gen-RecSys) extend beyond classical item ranking by producing open-ended content, which simultaneously unlocks richer user experiences and introduces new risks. On one hand, these systems can enhance personalization and appeal through dynamic explanations and multi-turn dialogues. On the other hand, they might venture into unknown territory-hallucinating nonexistent items, amplifying bias, or leaking private information. Traditional accuracy metrics cannot fully capture these challenges, as they fail to measure factual correctness, content safety, or alignment with user intent. This paper makes two main contributions. First, we categorize the evaluation challenges of Gen-RecSys into two groups: (i) existing concerns that are exacerbated by generative outputs (e.g., bias, privacy) and (ii) entirely new risks (e.g., item hallucinations, contradictory explanations). Second, we propose a holistic evaluation approach that includes scenario-based assessments and multi-metric checks-incorporating relevance, factual grounding, bias detection, and policy compliance. Our goal is to provide a guiding framework so researchers and practitioners can thoroughly assess Gen-RecSys, ensuring effective personalization and responsible deployment.
♻ ☆ Multi-Head RAG: Solving Multi-Aspect Problems with LLMs
Retrieval Augmented Generation (RAG) enhances the abilities of Large Language Models (LLMs) by enabling the retrieval of documents into the LLM context to provide more accurate and relevant responses. Existing RAG solutions do not focus on queries that may require fetching multiple documents with substantially different contents. Such queries occur frequently, but are challenging because the embeddings of these documents may be distant in the embedding space, making it hard to retrieve them all. This paper introduces Multi-Head RAG (MRAG), a novel scheme designed to address this gap with a simple yet powerful idea: leveraging activations of Transformer's multi-head attention layer, instead of the decoder layer, as keys for fetching multi-aspect documents. The driving observation is that different attention heads learn to capture different data aspects. Harnessing the corresponding activations results in embeddings that represent various facets of data items and queries, improving the retrieval accuracy for complex queries. We provide an evaluation methodology and metrics, multi-aspect datasets, and real-world use cases to demonstrate MRAG's effectiveness. We show MRAG's design advantages over 18 RAG baselines, empirical improvements of up to 20% in retrieval success ratios, and benefits for downstream LLM generation. MRAG can be seamlessly integrated with existing RAG frameworks and benchmarks.
♻ ☆ Affordable AI Assistants with Knowledge Graph of Thoughts
Large Language Models (LLMs) are revolutionizing the development of AI assistants capable of performing diverse tasks across domains. However, current state-of-the-art LLM-driven agents face significant challenges, including high operational costs and limited success rates on complex benchmarks like GAIA. To address these issues, we propose Knowledge Graph of Thoughts (KGoT), an innovative AI assistant architecture that integrates LLM reasoning with dynamically constructed knowledge graphs (KGs). KGoT extracts and structures task-relevant knowledge into a dynamic KG representation, iteratively enhanced through external tools such as math solvers, web crawlers, and Python scripts. Such structured representation of task-relevant knowledge enables low-cost models to solve complex tasks effectively while also minimizing bias and noise. For example, KGoT achieves a 29% improvement in task success rates on the GAIA benchmark compared to Hugging Face Agents with GPT-4o mini. Moreover, harnessing a smaller model dramatically reduces operational costs by over 36x compared to GPT-4o. Improvements for other models (e.g., Qwen2.5-32B and Deepseek-R1-70B) and benchmarks (e.g., SimpleQA) are similar. KGoT offers a scalable, affordable, versatile, and high-performing solution for AI assistants.
♻ ☆ USD: A User-Intent-Driven Sampling and Dual-Debiasing Framework for Large-Scale Homepage Recommendations
Large-scale homepage recommendations face critical challenges from pseudo-negative samples caused by exposure bias, where non-clicks may indicate inattention rather than disinterest. Existing work lacks thorough analysis of invalid exposures and typically addresses isolated aspects (e.g., sampling strategies), overlooking the critical impact of pseudo-positive samples - such as homepage clicks merely to visit marketing portals. We propose a unified framework for large-scale homepage recommendation sampling and debiasing. Our framework consists of two key components: (1) a user intent-aware negative sampling module to filter invalid exposure samples, and (2) an intent-driven dual-debiasing module that jointly corrects exposure bias and click bias. Extensive online experiments on Taobao demonstrate the efficacy of our framework, achieving significant improvements in user click-through rates (UCTR) by 35.4% and 14.5% in two variants of the marketing block on the Taobao homepage, Baiyibutie and Taobaomiaosha.
♻ ☆ Diffusion Augmented Retrieval: A Training-Free Approach to Interactive Text-to-Image Retrieval
Interactive Text-to-image retrieval (I-TIR) is an important enabler for a wide range of state-of-the-art services in domains such as e-commerce and education. However, current methods rely on finetuned Multimodal Large Language Models (MLLMs), which are costly to train and update, and exhibit poor generalizability. This latter issue is of particular concern, as: 1) finetuning narrows the pretrained distribution of MLLMs, thereby reducing generalizability; and 2) I-TIR introduces increasing query diversity and complexity. As a result, I-TIR solutions are highly likely to encounter queries and images not well represented in any training dataset. To address this, we propose leveraging Diffusion Models (DMs) for text-to-image mapping, to avoid finetuning MLLMs while preserving robust performance on complex queries. Specifically, we introduce Diffusion Augmented Retrieval (DAR), a framework that generates multiple intermediate representations via LLM-based dialogue refinements and DMs, producing a richer depiction of the user's information needs. This augmented representation facilitates more accurate identification of semantically and visually related images. Extensive experiments on four benchmarks show that for simple queries, DAR achieves results on par with finetuned I-TIR models, yet without incurring their tuning overhead. Moreover, as queries become more complex through additional conversational turns, DAR surpasses finetuned I-TIR models by up to 7.61% in Hits@10 after ten turns, illustrating its improved generalization for more intricate queries.
♻ ☆ U-Sticker: A Large-Scale Multi-Domain User Sticker Dataset for Retrieval and Personalization SIGIR'25
Instant messaging with texts and stickers has become a widely adopted communication medium, enabling efficient expression of user semantics and emotions. With the increased use of stickers conveying information and feelings, sticker retrieval and recommendation has emerged as an important area of research. However, a major limitation in existing literature has been the lack of datasets capturing temporal and user-specific sticker interactions, which has hindered further progress in user modeling and sticker personalization. To address this, we introduce User-Sticker, a dataset that includes temporal and user anonymous ID across conversations. It is the largest publicly available sticker dataset to date, containing 22K unique users, 370K stickers, and 8.3M messages. The raw data was collected from a popular messaging platform from 67 conversations over 720 hours of crawling. All text and image data were carefully vetted for safety and privacy checks and modifications. Spanning 10 domains, the U-Sticker dataset captures rich temporal, multilingual, and cross-domain behaviors not previously available in other datasets. Extensive quantitative and qualitative experiments demonstrate U-Sticker's practical applications in user behavior modeling and personalized recommendation and highlight its potential to further research areas in personalized retrieval and conversational studies. U-Sticker dataset is publicly available.
comment: Accepted at SIGIR'25
♻ ☆ Shifting from Ranking to Set Selection for Retrieval Augmented Generation ACL 2025
Retrieval in Retrieval-Augmented Generation(RAG) must ensure that retrieved passages are not only individually relevant but also collectively form a comprehensive set. Existing approaches primarily rerank top-k passages based on their individual relevance, often failing to meet the information needs of complex queries in multi-hop question answering. In this work, we propose a set-wise passage selection approach and introduce SETR, which explicitly identifies the information requirements of a query through Chain-of-Thought reasoning and selects an optimal set of passages that collectively satisfy those requirements. Experiments on multi-hop RAG benchmarks show that SETR outperforms both proprietary LLM-based rerankers and open-source baselines in terms of answer correctness and retrieval quality, providing an effective and efficient alternative to traditional rerankers in RAG systems. The code is available at https://github.com/LGAI-Research/SetR
comment: Accepted to ACL 2025 main (Oral Presentation)
Machine Learning 154
☆ Impact of Pretraining Word Co-occurrence on Compositional Generalization in Multimodal Models
CLIP and large multimodal models (LMMs) have better accuracy on examples involving concepts that are highly represented in the training data. However, the role of concept combinations in the training data on compositional generalization is largely unclear -- for instance, how does accuracy vary when a common object appears in an uncommon pairing with another object? In this paper, we investigate how word co-occurrence statistics in the pretraining dataset (a proxy for co-occurrence of visual concepts) impacts CLIP/LMM performance. To disentangle the effects of word co-occurrence frequencies from single-word frequencies, we measure co-occurrence with pointwise mutual information (PMI), which normalizes the joint probability of two words co-occurring by the probability of co-occurring independently. Using synthetically generated images with a variety of concept pairs, we show a strong correlation between PMI in the CLIP pretraining data and zero-shot accuracy in CLIP models trained on LAION-400M (r=0.97 and 14% accuracy gap between images in the top and bottom 5% of PMI values), demonstrating that even accuracy on common concepts is affected by the combination of concepts in the image. Leveraging this finding, we reproduce this effect in natural images by editing them to contain pairs with varying PMI, resulting in a correlation of r=0.75. Finally, we demonstrate that this behavior in CLIP transfers to LMMs built on top of CLIP (r=0.70 for TextVQA, r=0.62 for VQAv2). Our findings highlight the need for algorithms and architectures that improve compositional generalization in multimodal models without scaling the training data combinatorially. Our code is available at https://github.com/helenqu/multimodal-pretraining-pmi.
☆ Single-pass Adaptive Image Tokenization for Minimum Program Search
According to Algorithmic Information Theory (AIT) -- Intelligent representations compress data into the shortest possible program that can reconstruct its content, exhibiting low Kolmogorov Complexity (KC). In contrast, most visual representation learning systems use fixed-length representations for all inputs, ignoring variations in complexity or familiarity. Recent adaptive tokenization methods address this by allocating variable-length representations but typically require test-time search over multiple encodings to find the most predictive one. Inspired by Kolmogorov Complexity principles, we propose a single-pass adaptive tokenizer, KARL, which predicts the appropriate number of tokens for an image in a single forward pass, halting once its approximate KC is reached. The token count serves as a proxy for the minimum description length. KARL's training procedure closely resembles the Upside-Down Reinforcement Learning paradigm, as it learns to conditionally predict token halting based on a desired reconstruction quality. KARL matches the performance of recent adaptive tokenizers while operating in a single pass. We present scaling laws for KARL, analyzing the role of encoder/decoder size, continuous vs. discrete tokenization and more. Additionally, we offer a conceptual study drawing an analogy between Adaptive Image Tokenization and Algorithmic Information Theory, examining the predicted image complexity (KC) across axes such as structure vs. noise and in- vs. out-of-distribution familiarity -- revealing alignment with human intuition.
comment: Code at: https://github.com/ShivamDuggal4/karl Keywords: Representation Learning, Adaptive Tokenization, Compression, Algorithmic Information Theory, Kolmogorov Complexity, Upside-Down RL
☆ Skip a Layer or Loop it? Test-Time Depth Adaptation of Pretrained LLMs
Can a pretrained neural network adapt its architecture to different inputs without any finetuning? Do we need all layers for simple tasks, and are they adequate for challenging tasks? We found that the layers of a pretrained large language model (LLM) can be manipulated as separate modules to build a better and even shallower model customized for each test sample. In particular, each layer from the pretrained model can be skipped/pruned or repeated multiple times as recurrent neural networks (RNN), and stacked with others in arbitrary orders, yielding a chain-of-layers (CoLa) per sample. This compositional space greatly expands the scope of existing works on looped/recurrent pretrained modules, layer pruning, or early-exit networks. We develop a Monte Carlo Tree Search (MCTS) protocol to explore and identify the optimal CoLa for each sample from math and commonsense reasoning benchmarks. Compared to a static model of a fixed depth, CoLa allows shortcut paths (fast thinking), recurrence of the same layer(s) (slow thinking), and combining both, offering more flexible, dynamic architectures for different inputs. We conduct an extensive analysis of the MCTS-optimized CoLa, which leads to two key findings: (1) For >75% of samples with correct predictions by the original LLM, we can find shorter CoLa, suggesting a large space for improving inference efficiency; (2) For >60% of samples with originally incorrect predictions, we can identify CoLa achieving correct predictions, suggesting a large space of performance enhancement. Our results highlight the shortcomings of using a fixed architecture of pre-trained LLMs for inference on different samples and pave the way to unlock the generalization power of test-time depth adaptation.
comment: 9 pages, 7 figures
☆ EXPO: Stable Reinforcement Learning with Expressive Policies
We study the problem of training and fine-tuning expressive policies with online reinforcement learning (RL) given an offline dataset. Training expressive policy classes with online RL present a unique challenge of stable value maximization. Unlike simpler Gaussian policies commonly used in online RL, expressive policies like diffusion and flow-matching policies are parameterized by a long denoising chain, which hinders stable gradient propagation from actions to policy parameters when optimizing against some value function. Our key insight is that we can address stable value maximization by avoiding direct optimization over value with the expressive policy and instead construct an on-the-fly RL policy to maximize Q-value. We propose Expressive Policy Optimization (EXPO), a sample-efficient online RL algorithm that utilizes an on-the-fly policy to maximize value with two parameterized policies -- a larger expressive base policy trained with a stable imitation learning objective and a light-weight Gaussian edit policy that edits the actions sampled from the base policy toward a higher value distribution. The on-the-fly policy optimizes the actions from the base policy with the learned edit policy and chooses the value maximizing action from the base and edited actions for both sampling and temporal-difference (TD) backup. Our approach yields up to 2-3x improvement in sample efficiency on average over prior methods both in the setting of fine-tuning a pretrained policy given offline data and in leveraging offline data to train online.
☆ Why is Your Language Model a Poor Implicit Reward Model?
Reward models are key to language model post-training and inference pipelines. Conveniently, recent work showed that every language model defines an implicit reward model (IM-RM), without requiring any architectural changes. However, such IM-RMs tend to generalize worse, especially out-of-distribution, compared to explicit reward models (EX-RMs) that apply a dedicated linear head over the hidden representations of a language model. The existence of a generalization gap is puzzling, as EX-RMs and IM-RMs are nearly identical. They can be trained using the same data, loss function, and language model, and differ only in how the reward is computed. Towards a fundamental understanding of the implicit biases underlying different reward model types, we investigate the root cause of this gap. Our main finding, backed by theory and experiments, is that IM-RMs rely more heavily on superficial token-level cues. Consequently, they often generalize worse than EX-RMs under token-level distribution shifts, as well as in-distribution. Furthermore, we provide evidence against alternative hypotheses for the generalization gap. Most notably, we challenge the intuitive claim that IM-RMs struggle in tasks where generation is harder than verification because they can operate both as a verifier and a generator. Taken together, our results highlight that seemingly minor design choices can substantially impact the generalization behavior of reward models.
☆ Reinforcement Learning with Action Chunking
We present Q-chunking, a simple yet effective recipe for improving reinforcement learning (RL) algorithms for long-horizon, sparse-reward tasks. Our recipe is designed for the offline-to-online RL setting, where the goal is to leverage an offline prior dataset to maximize the sample-efficiency of online learning. Effective exploration and sample-efficient learning remain central challenges in this setting, as it is not obvious how the offline data should be utilized to acquire a good exploratory policy. Our key insight is that action chunking, a technique popularized in imitation learning where sequences of future actions are predicted rather than a single action at each timestep, can be applied to temporal difference (TD)-based RL methods to mitigate the exploration challenge. Q-chunking adopts action chunking by directly running RL in a 'chunked' action space, enabling the agent to (1) leverage temporally consistent behaviors from offline data for more effective online exploration and (2) use unbiased $n$-step backups for more stable and efficient TD learning. Our experimental results demonstrate that Q-chunking exhibits strong offline performance and online sample efficiency, outperforming prior best offline-to-online methods on a range of long-horizon, sparse-reward manipulation tasks.
comment: 25 pages, 15 figures
☆ Prospective Learning in Retrospect
In most real-world applications of artificial intelligence, the distributions of the data and the goals of the learners tend to change over time. The Probably Approximately Correct (PAC) learning framework, which underpins most machine learning algorithms, fails to account for dynamic data distributions and evolving objectives, often resulting in suboptimal performance. Prospective learning is a recently introduced mathematical framework that overcomes some of these limitations. We build on this framework to present preliminary results that improve the algorithm and numerical results, and extend prospective learning to sequential decision-making scenarios, specifically foraging. Code is available at: https://github.com/neurodata/prolearn2.
comment: Accepted to AGI 2025
☆ Dynamic Chunking for End-to-End Hierarchical Sequence Modeling
Despite incredible progress in language models (LMs) in recent years, largely resulting from moving away from specialized models designed for specific tasks to general models based on powerful architectures (e.g. the Transformer) that learn everything from raw data, pre-processing steps such as tokenization remain a barrier to true end-to-end foundation models. We introduce a collection of new techniques that enable a dynamic chunking mechanism which automatically learns content -- and context -- dependent segmentation strategies learned jointly with the rest of the model. Incorporating this into an explicit hierarchical network (H-Net) allows replacing the (implicitly hierarchical) tokenization-LM-detokenization pipeline with a single model learned fully end-to-end. When compute- and data- matched, an H-Net with one stage of hierarchy operating at the byte level outperforms a strong Transformer language model operating over BPE tokens. Iterating the hierarchy to multiple stages further increases its performance by modeling multiple levels of abstraction, demonstrating significantly better scaling with data and matching a token-based Transformer of twice its size. H-Nets pretrained on English show significantly increased character-level robustness, and qualitatively learn meaningful data-dependent chunking strategies without any heuristics or explicit supervision. Finally, the H-Net's improvement over tokenized pipelines is further increased in languages and modalities with weaker tokenization heuristics, such as Chinese and code, or DNA sequences (nearly 4x improvement in data efficiency over baselines), showing the potential of true end-to-end models that learn and scale better from unprocessed data.
☆ TinierHAR: Towards Ultra-Lightweight Deep Learning Models for Efficient Human Activity Recognition on Edge Devices
Human Activity Recognition (HAR) on resource-constrained wearable devices demands inference models that harmonize accuracy with computational efficiency. This paper introduces TinierHAR, an ultra-lightweight deep learning architecture that synergizes residual depthwise separable convolutions, gated recurrent units (GRUs), and temporal aggregation to achieve SOTA efficiency without compromising performance. Evaluated across 14 public HAR datasets, TinierHAR reduces Parameters by 2.7x (vs. TinyHAR) and 43.3x (vs. DeepConvLSTM), and MACs by 6.4x and 58.6x, respectively, while maintaining the averaged F1-scores. Beyond quantitative gains, this work provides the first systematic ablation study dissecting the contributions of spatial-temporal components across proposed TinierHAR, prior SOTA TinyHAR, and the classical DeepConvLSTM, offering actionable insights for designing efficient HAR systems. We finally discussed the findings and suggested principled design guidelines for future efficient HAR. To catalyze edge-HAR research, we open-source all materials in this work for future benchmarking\footnote{https://github.com/zhaxidele/TinierHAR}
☆ Low Resource Reconstruction Attacks Through Benign Prompts
The recent advances in generative models such as diffusion models have raised several risks and concerns related to privacy, copyright infringements and data stewardship. To better understand and control the risks, various researchers have created techniques, experiments and attacks that reconstruct images, or part of images, from the training set. While these techniques already establish that data from the training set can be reconstructed, they often rely on high-resources, excess to the training set as well as well-engineered and designed prompts. In this work, we devise a new attack that requires low resources, assumes little to no access to the actual training set, and identifies, seemingly, benign prompts that lead to potentially-risky image reconstruction. This highlights the risk that images might even be reconstructed by an uninformed user and unintentionally. For example, we identified that, with regard to one existing model, the prompt ``blue Unisex T-Shirt'' can generate the face of a real-life human model. Our method builds on an intuition from previous works which leverages domain knowledge and identifies a fundamental vulnerability that stems from the use of scraped data from e-commerce platforms, where templated layouts and images are tied to pattern-like prompts.
☆ Towards Continuous Home Cage Monitoring: An Evaluation of Tracking and Identification Strategies for Laboratory Mice
Continuous, automated monitoring of laboratory mice enables more accurate data collection and improves animal welfare through real-time insights. Researchers can achieve a more dynamic and clinically relevant characterization of disease progression and therapeutic effects by integrating behavioral and physiological monitoring in the home cage. However, providing individual mouse metrics is difficult because of their housing density, similar appearances, high mobility, and frequent interactions. To address these challenges, we develop a real-time identification (ID) algorithm that accurately assigns ID predictions to mice wearing custom ear tags in digital home cages monitored by cameras. Our pipeline consists of three parts: (1) a custom multiple object tracker (MouseTracks) that combines appearance and motion cues from mice; (2) a transformer-based ID classifier (Mouseformer); and (3) a tracklet associator linear program to assign final ID predictions to tracklets (MouseMap). Our models assign an animal ID based on custom ear tags at 30 frames per second with 24/7 cage coverage. We show that our custom tracking and ID pipeline improves tracking efficiency and lowers ID switches across mouse strains and various environmental factors compared to current mouse tracking methods.
☆ Plausible Counterfactual Explanations of Recommendations
Explanations play a variety of roles in various recommender systems, from a legally mandated afterthought, through an integral element of user experience, to a key to persuasiveness. A natural and useful form of an explanation is the Counterfactual Explanation (CE). We present a method for generating highly plausible CEs in recommender systems and evaluate it both numerically and with a user study.
comment: 8 pages, 3 figures, 6 tables
☆ A statistical physics framework for optimal learning
Learning is a complex dynamical process shaped by a range of interconnected decisions. Careful design of hyperparameter schedules for artificial neural networks or efficient allocation of cognitive resources by biological learners can dramatically affect performance. Yet, theoretical understanding of optimal learning strategies remains sparse, especially due to the intricate interplay between evolving meta-parameters and nonlinear learning dynamics. The search for optimal protocols is further hindered by the high dimensionality of the learning space, often resulting in predominantly heuristic, difficult to interpret, and computationally demanding solutions. Here, we combine statistical physics with control theory in a unified theoretical framework to identify optimal protocols in prototypical neural network models. In the high-dimensional limit, we derive closed-form ordinary differential equations that track online stochastic gradient descent through low-dimensional order parameters. We formulate the design of learning protocols as an optimal control problem directly on the dynamics of the order parameters with the goal of minimizing the generalization error at the end of training. This framework encompasses a variety of learning scenarios, optimization constraints, and control budgets. We apply it to representative cases, including optimal curricula, adaptive dropout regularization and noise schedules in denoising autoencoders. We find nontrivial yet interpretable strategies highlighting how optimal protocols mediate crucial learning tradeoffs, such as maximizing alignment with informative input directions while minimizing noise fitting. Finally, we show how to apply our framework to real datasets. Our results establish a principled foundation for understanding and designing optimal learning protocols and suggest a path toward a theory of meta-learning grounded in statistical physics.
comment: 35 pages, 13 figures
☆ Agentic Retrieval of Topics and Insights from Earnings Calls SIGIR
Tracking the strategic focus of companies through topics in their earnings calls is a key task in financial analysis. However, as industries evolve, traditional topic modeling techniques struggle to dynamically capture emerging topics and their relationships. In this work, we propose an LLM-agent driven approach to discover and retrieve emerging topics from quarterly earnings calls. We propose an LLM-agent to extract topics from documents, structure them into a hierarchical ontology, and establish relationships between new and existing topics through a topic ontology. We demonstrate the use of extracted topics to infer company-level insights and emerging trends over time. We evaluate our approach by measuring ontology coherence, topic evolution accuracy, and its ability to surface emerging financial trends.
comment: The 2nd Workshop on Financial Information Retrieval in the Era of Generative AI, The 48th International ACM SIGIR Conference on Research and Development in Information Retrieval July 13-17, 2025 | Padua, Italy
☆ Efficient Causal Discovery for Autoregressive Time Series
In this study, we present a novel constraint-based algorithm for causal structure learning specifically designed for nonlinear autoregressive time series. Our algorithm significantly reduces computational complexity compared to existing methods, making it more efficient and scalable to larger problems. We rigorously evaluate its performance on synthetic datasets, demonstrating that our algorithm not only outperforms current techniques, but also excels in scenarios with limited data availability. These results highlight its potential for practical applications in fields requiring efficient and accurate causal inference from nonlinear time series data.
comment: 10 pages, 8 figures
☆ UnIT: Scalable Unstructured Inference-Time Pruning for MAC-efficient Neural Inference on MCUs
Existing pruning methods are typically applied during training or compile time and often rely on structured sparsity. While compatible with low-power microcontrollers (MCUs), structured pruning underutilizes the opportunity for fine-grained efficiency on devices without SIMD support or parallel compute. To address these limitations, we introduce UnIT (Unstructured Inference-Time pruning), a lightweight method that dynamically identifies and skips unnecessary multiply-accumulate (MAC) operations during inference, guided by input-specific activation patterns. Unlike structured pruning, UnIT embraces irregular sparsity and does not require retraining or hardware specialization. It transforms pruning decisions into lightweight comparisons, replacing multiplications with threshold checks and approximated divisions. UnIT further optimizes compute by reusing threshold computations across multiple connections and applying layer- and group-specific pruning sensitivity. We present three fast, hardware-friendly division approximations tailored to the capabilities of common embedded platforms. Demonstrated on the MSP430 microcontroller, UnIT achieves 11.02% to 82.03% MAC reduction, 27.30% to 84.19% faster inference, and 27.33% to 84.38% lower energy consumption compared to training-time pruned models, while maintaining accuracy with 0.48-7%. Under domain shift, UnIT matches or exceeds the accuracy of retrained models while requiring significantly fewer MACs. These results establish unstructured inference-time pruning as a viable and practical solution for efficient, retraining-free deployment of deep neural networks on MCUs.
comment: Submitted to SenSys 2026 on July 1, 2025
☆ SAMO: A Lightweight Sharpness-Aware Approach for Multi-Task Optimization with Joint Global-Local Perturbation
Multi-task learning (MTL) enables a joint model to capture commonalities across multiple tasks, reducing computation costs and improving data efficiency. However, a major challenge in MTL optimization is task conflicts, where the task gradients differ in direction or magnitude, limiting model performance compared to single-task counterparts. Sharpness-aware minimization (SAM) minimizes task loss while simultaneously reducing the sharpness of the loss landscape. Our empirical observations show that SAM effectively mitigates task conflicts in MTL. Motivated by these findings, we explore integrating SAM into MTL but face two key challenges. While both the average loss gradient and individual task gradients-referred to as global and local information-contribute to SAM, how to combine them remains unclear. Moreover, directly computing each task gradient introduces significant computational and memory overheads. To address these challenges, we propose SAMO, a lightweight \textbf{S}harpness-\textbf{A}ware \textbf{M}ulti-task \textbf{O}ptimization approach, that leverages a joint global-local perturbation. The local perturbations are approximated using only forward passes and are layerwise normalized to improve efficiency. Extensive experiments on a suite of multi-task benchmarks demonstrate both the effectiveness and efficiency of our method. Code is available at https://github.com/OptMN-Lab/SAMO.
☆ Can AI-predicted complexes teach machine learning to compute drug binding affinity?
We evaluate the feasibility of using co-folding models for synthetic data augmentation in training machine learning-based scoring functions (MLSFs) for binding affinity prediction. Our results show that performance gains depend critically on the structural quality of augmented data. In light of this, we established simple heuristics for identifying high-quality co-folding predictions without reference structures, enabling them to substitute for experimental structures in MLSF training. Our study informs future data augmentation strategies based on co-folding models.
☆ Edge-ASR: Towards Low-Bit Quantization of Automatic Speech Recognition Models
Recent advances in Automatic Speech Recognition (ASR) have demonstrated remarkable accuracy and robustness in diverse audio applications, such as live transcription and voice command processing. However, deploying these models on resource constrained edge devices (e.g., IoT device, wearables) still presents substantial challenges due to strict limits on memory, compute and power. Quantization, particularly Post-Training Quantization (PTQ), offers an effective way to reduce model size and inference cost without retraining. Despite its importance, the performance implications of various advanced quantization methods and bit-width configurations on ASR models remain unclear. In this work, we present a comprehensive benchmark of eight state-of-the-art (SOTA) PTQ methods applied to two leading edge-ASR model families, Whisper and Moonshine. We systematically evaluate model performances (i.e., accuracy, memory I/O and bit operations) across seven diverse datasets from the open ASR leaderboard, analyzing the impact of quantization and various configurations on both weights and activations. Built on an extension of the LLM compression toolkit, our framework integrates edge-ASR models, diverse advanced quantization algorithms, a unified calibration and evaluation data pipeline, and detailed analysis tools. Our results characterize the trade-offs between efficiency and accuracy, demonstrating that even 3-bit quantization can succeed on high capacity models when using advanced PTQ techniques. These findings provide valuable insights for optimizing ASR models on low-power, always-on edge devices.
☆ Improving AEBS Validation Through Objective Intervention Classification Leveraging the Prediction Divergence Principle
The safety validation of automatic emergency braking system (AEBS) requires accurately distinguishing between false positive (FP) and true positive (TP) system activations. While simulations allow straightforward differentiation by comparing scenarios with and without interventions, analyzing activations from open-loop resimulations - such as those from field operational testing (FOT) - is more complex. This complexity arises from scenario parameter uncertainty and the influence of driver interventions in the recorded data. Human labeling is frequently used to address these challenges, relying on subjective assessments of intervention necessity or situational criticality, potentially introducing biases and limitations. This work proposes a rule-based classification approach leveraging the Prediction Divergence Principle (PDP) to address those issues. Applied to a simplified AEBS, the proposed method reveals key strengths, limitations, and system requirements for effective implementation. The findings suggest that combining this approach with human labeling may enhance the transparency and consistency of classification, thereby improving the overall validation process. While the rule set for classification derived in this work adopts a conservative approach, the paper outlines future directions for refinement and broader applicability. Finally, this work highlights the potential of such methods to complement existing practices, paving the way for more reliable and reproducible AEBS validation frameworks.
comment: This work has been accepted for publication at the 2025 IEEE International Automated Vehicle Validation Conference (IAVVC)
☆ Mitigating Watermark Stealing Attacks in Generative Models via Multi-Key Watermarking
Watermarking offers a promising solution for GenAI providers to establish the provenance of their generated content. A watermark is a hidden signal embedded in the generated content, whose presence can later be verified using a secret watermarking key. A threat to GenAI providers are \emph{watermark stealing} attacks, where users forge a watermark into content that was \emph{not} generated by the provider's models without access to the secret key, e.g., to falsely accuse the provider. Stealing attacks collect \emph{harmless} watermarked samples from the provider's model and aim to maximize the expected success rate of generating \emph{harmful} watermarked samples. Our work focuses on mitigating stealing attacks while treating the underlying watermark as a black-box. Our contributions are: (i) Proposing a multi-key extension to mitigate stealing attacks that can be applied post-hoc to any watermarking method across any modality. (ii) We provide theoretical guarantees and demonstrate empirically that our method makes forging substantially less effective across multiple datasets, and (iii) we formally define the threat of watermark forging as the task of generating harmful, watermarked content and model this threat via security games.
☆ Re-Bottleneck: Latent Re-Structuring for Neural Audio Autoencoders SP 2025
Neural audio codecs and autoencoders have emerged as versatile models for audio compression, transmission, feature-extraction, and latent-space generation. However, a key limitation is that most are trained to maximize reconstruction fidelity, often neglecting the specific latent structure necessary for optimal performance in diverse downstream applications. We propose a simple, post-hoc framework to address this by modifying the bottleneck of a pre-trained autoencoder. Our method introduces a "Re-Bottleneck", an inner bottleneck trained exclusively through latent space losses to instill user-defined structure. We demonstrate the framework's effectiveness in three experiments. First, we enforce an ordering on latent channels without sacrificing reconstruction quality. Second, we align latents with semantic embeddings, analyzing the impact on downstream diffusion modeling. Third, we introduce equivariance, ensuring that a filtering operation on the input waveform directly corresponds to a specific transformation in the latent space. Ultimately, our Re-Bottleneck framework offers a flexible and efficient way to tailor representations of neural audio models, enabling them to seamlessly meet the varied demands of different applications with minimal additional training.
comment: Accepted at IEEE MLSP 2025
☆ Predicting and generating antibiotics against future pathogens with ApexOracle
Antimicrobial resistance (AMR) is escalating and outpacing current antibiotic development. Thus, discovering antibiotics effective against emerging pathogens is becoming increasingly critical. However, existing approaches cannot rapidly identify effective molecules against novel pathogens or emerging drug-resistant strains. Here, we introduce ApexOracle, an artificial intelligence (AI) model that both predicts the antibacterial potency of existing compounds and designs de novo molecules active against strains it has never encountered. Departing from models that rely solely on molecular features, ApexOracle incorporates pathogen-specific context through the integration of molecular features captured via a foundational discrete diffusion language model and a dual-embedding framework that combines genomic- and literature-derived strain representations. Across diverse bacterial species and chemical modalities, ApexOracle consistently outperformed state-of-the-art approaches in activity prediction and demonstrated reliable transferability to novel pathogens with little or no antimicrobial data. Its unified representation-generation architecture further enables the in silico creation of "new-to-nature" molecules with high predicted efficacy against priority threats. By pairing rapid activity prediction with targeted molecular generation, ApexOracle offers a scalable strategy for countering AMR and preparing for future infectious-disease outbreaks.
comment: 3 figures
☆ Principled Foundations for Preference Optimization
In this paper, we show that direct preference optimization (DPO) is a very specific form of a connection between two major theories in the ML context of learning from preferences: loss functions (Savage) and stochastic choice (Doignon-Falmagne and Machina). The connection is established for all of Savage's losses and at this level of generality, (i) it includes support for abstention on the choice theory side, (ii) it includes support for non-convex objectives on the ML side, and (iii) it allows to frame for free some notable extensions of the DPO setting, including margins and corrections for length. Getting to understand how DPO operates from a general principled perspective is crucial because of the huge and diverse application landscape of models, because of the current momentum around DPO, but also -- and importantly -- because many state of the art variations on DPO definitely occupy a small region of the map that we cover. It also helps to understand the pitfalls of departing from this map, and figure out workarounds.
☆ Credit Risk Analysis for SMEs Using Graph Neural Networks in Supply Chain
Small and Medium-sized Enterprises (SMEs) are vital to the modern economy, yet their credit risk analysis often struggles with scarce data, especially for online lenders lacking direct credit records. This paper introduces a Graph Neural Network (GNN)-based framework, leveraging SME interactions from transaction and social data to map spatial dependencies and predict loan default risks. Tests on real-world datasets from Discover and Ant Credit (23.4M nodes for supply chain analysis, 8.6M for default prediction) show the GNN surpasses traditional and other GNN baselines, with AUCs of 0.995 and 0.701 for supply chain mining and default prediction, respectively. It also helps regulators model supply chain disruption impacts on banks, accurately forecasting loan defaults from material shortages, and offers Federal Reserve stress testers key data for CCAR risk buffers. This approach provides a scalable, effective tool for assessing SME credit risk.
comment: The paper will be published on 2025 International Conference on Big Data, Artificial Intelligence and Digital Economy
☆ Optimization Guarantees for Square-Root Natural-Gradient Variational Inference
Variational inference with natural-gradient descent often shows fast convergence in practice, but its theoretical convergence guarantees have been challenging to establish. This is true even for the simplest cases that involve concave log-likelihoods and use a Gaussian approximation. We show that the challenge can be circumvented for such cases using a square-root parameterization for the Gaussian covariance. This approach establishes novel convergence guarantees for natural-gradient variational-Gaussian inference and its continuous-time gradient flow. Our experiments demonstrate the effectiveness of natural gradient methods and highlight their advantages over algorithms that use Euclidean or Wasserstein geometries.
Pre-Trained AI Model Assisted Online Decision-Making under Missing Covariates: A Theoretical Perspective
We study a sequential contextual decision-making problem in which certain covariates are missing but can be imputed using a pre-trained AI model. From a theoretical perspective, we analyze how the presence of such a model influences the regret of the decision-making process. We introduce a novel notion called "model elasticity", which quantifies the sensitivity of the reward function to the discrepancy between the true covariate and its imputed counterpart. This concept provides a unified way to characterize the regret incurred due to model imputation, regardless of the underlying missingness mechanism. More surprisingly, we show that under the missing at random (MAR) setting, it is possible to sequentially calibrate the pre-trained model using tools from orthogonal statistical learning and doubly robust regression. This calibration significantly improves the quality of the imputed covariates, leading to much better regret guarantees. Our analysis highlights the practical value of having an accurate pre-trained model in sequential decision-making tasks and suggests that model elasticity may serve as a fundamental metric for understanding and improving the integration of pre-trained models in a wide range of data-driven decision-making problems.
☆ "So, Tell Me About Your Policy...": Distillation of interpretable policies from Deep Reinforcement Learning agents
Recent advances in Reinforcement Learning (RL) largely benefit from the inclusion of Deep Neural Networks, boosting the number of novel approaches proposed in the field of Deep Reinforcement Learning (DRL). These techniques demonstrate the ability to tackle complex games such as Atari, Go, and other real-world applications, including financial trading. Nevertheless, a significant challenge emerges from the lack of interpretability, particularly when attempting to comprehend the underlying patterns learned, the relative importance of the state features, and how they are integrated to generate the policy's output. For this reason, in mission-critical and real-world settings, it is often preferred to deploy a simpler and more interpretable algorithm, although at the cost of performance. In this paper, we propose a novel algorithm, supported by theoretical guarantees, that can extract an interpretable policy (e.g., a linear policy) without disregarding the peculiarities of expert behavior. This result is obtained by considering the advantage function, which includes information about why an action is superior to the others. In contrast to previous works, our approach enables the training of an interpretable policy using previously collected experience. The proposed algorithm is empirically evaluated on classic control environments and on a financial trading scenario, demonstrating its ability to extract meaningful information from complex expert policies.
☆ Towards Benchmarking Foundation Models for Tabular Data With Text ICML 2025
Foundation models for tabular data are rapidly evolving, with increasing interest in extending them to support additional modalities such as free-text features. However, existing benchmarks for tabular data rarely include textual columns, and identifying real-world tabular datasets with semantically rich text features is non-trivial. We propose a series of simple yet effective ablation-style strategies for incorporating text into conventional tabular pipelines. Moreover, we benchmark how state-of-the-art tabular foundation models can handle textual data by manually curating a collection of real-world tabular datasets with meaningful textual features. Our study is an important step towards improving benchmarking of foundation models for tabular data with text.
comment: Accepted at Foundation Models for Structured Data workshop at ICML 2025
☆ An Empirical Bernstein Inequality for Dependent Data in Hilbert Spaces and Applications
Learning from non-independent and non-identically distributed data poses a persistent challenge in statistical learning. In this study, we introduce data-dependent Bernstein inequalities tailored for vector-valued processes in Hilbert space. Our inequalities apply to both stationary and non-stationary processes and exploit the potential rapid decay of correlations between temporally separated variables to improve estimation. We demonstrate the utility of these bounds by applying them to covariance operator estimation in the Hilbert-Schmidt norm and to operator learning in dynamical systems, achieving novel risk bounds. Finally, we perform numerical experiments to illustrate the practical implications of these bounds in both contexts.
comment: In The 28th International Conference on Artificial Intelligence and Statistics (2025)
☆ On the Effect of Instruction Tuning Loss on Generalization ACL
Instruction Tuning has emerged as a pivotal post-training paradigm that enables pre-trained language models to better follow user instructions. Despite its significance, little attention has been given to optimizing the loss function used. A fundamental, yet often overlooked, question is whether the conventional auto-regressive objective - where loss is computed only on response tokens, excluding prompt tokens - is truly optimal for instruction tuning. In this work, we systematically investigate the impact of differentially weighting prompt and response tokens in instruction tuning loss, and propose Weighted Instruction Tuning (WIT) as a better alternative to conventional instruction tuning. Through extensive experiments on five language models of different families and scale, three finetuning datasets of different sizes, and five diverse evaluation benchmarks, we show that the standard instruction tuning loss often yields suboptimal performance and limited robustness to input prompt variations. We find that a low-to-moderate weight for prompt tokens coupled with a moderate-to-high weight for response tokens yields the best-performing models across settings and also serve as better starting points for the subsequent preference alignment training. These findings highlight the need to reconsider instruction tuning loss and offer actionable insights for developing more robust and generalizable models. Our code is open-sourced at https://github.com/kowndinya-renduchintala/WIT.
comment: Transactions of the Association for Computational Linguistics (TACL)
☆ Pay Attention to Attention Distribution: A New Local Lipschitz Bound for Transformers
We present a novel local Lipschitz bound for self-attention blocks of transformers. This bound is based on a refined closed-form expression for the spectral norm of the softmax function. The resulting bound is not only more accurate than in the prior art, but also unveils the dependence of the Lipschitz constant on attention score maps. Based on the new findings, we suggest an explanation of the way distributions inside the attention map affect the robustness from the Lipschitz constant perspective. We also introduce a new lightweight regularization term called JaSMin (Jacobian Softmax norm Minimization), which boosts the transformer's robustness and decreases local Lipschitz constants of the whole network.
☆ Deep Survival Analysis in Multimodal Medical Data: A Parametric and Probabilistic Approach with Competing Risks
Accurate survival prediction is critical in oncology for prognosis and treatment planning. Traditional approaches often rely on a single data modality, limiting their ability to capture the complexity of tumor biology. To address this challenge, we introduce a multimodal deep learning framework for survival analysis capable of modeling both single and competing risks scenarios, evaluating the impact of integrating multiple medical data sources on survival predictions. We propose SAMVAE (Survival Analysis Multimodal Variational Autoencoder), a novel deep learning architecture designed for survival prediction that integrates six data modalities: clinical variables, four molecular profiles, and histopathological images. SAMVAE leverages modality specific encoders to project inputs into a shared latent space, enabling robust survival prediction while preserving modality specific information. Its parametric formulation enables the derivation of clinically meaningful statistics from the output distributions, providing patient-specific insights through interactive multimedia that contribute to more informed clinical decision-making and establish a foundation for interpretable, data-driven survival analysis in oncology. We evaluate SAMVAE on two cancer cohorts breast cancer and lower grade glioma applying tailored preprocessing, dimensionality reduction, and hyperparameter optimization. The results demonstrate the successful integration of multimodal data for both standard survival analysis and competing risks scenarios across different datasets. Our model achieves competitive performance compared to state-of-the-art multimodal survival models. Notably, this is the first parametric multimodal deep learning architecture to incorporate competing risks while modeling continuous time to a specific event, using both tabular and image data.
comment: 29 pages, 9 Figures
☆ Space-Filling Regularization for Robust and Interpretable Nonlinear State Space Models
The state space dynamics representation is the most general approach for nonlinear systems and often chosen for system identification. During training, the state trajectory can deform significantly leading to poor data coverage of the state space. This can cause significant issues for space-oriented training algorithms which e.g. rely on grid structures, tree partitioning, or similar. Besides hindering training, significant state trajectory deformations also deteriorate interpretability and robustness properties. This paper proposes a new type of space-filling regularization that ensures a favorable data distribution in state space via introducing a data-distribution-based penalty. This method is demonstrated in local model network architectures where good interpretability is a major concern. The proposed approach integrates ideas from modeling and design of experiments for state space structures. This is why we present two regularization techniques for the data point distributions of the state trajectories for local affine state space models. Beyond that, we demonstrate the results on a widely known system identification benchmark.
☆ Approximation Depth of Convex Polytopes
We study approximations of polytopes in the standard model for computing polytopes using Minkowski sums and (convex hulls of) unions. Specifically, we study the ability to approximate a target polytope by polytopes of a given depth. Our main results imply that simplices can only be ``trivially approximated''. On the way, we obtain a characterization of simplices as the only ``outer additive'' convex bodies.
☆ Synchronizing Task Behavior: Aligning Multiple Tasks during Test-Time Training ICCV 2025
Generalizing neural networks to unseen target domains is a significant challenge in real-world deployments. Test-time training (TTT) addresses this by using an auxiliary self-supervised task to reduce the domain gap caused by distribution shifts between the source and target. However, we find that when models are required to perform multiple tasks under domain shifts, conventional TTT methods suffer from unsynchronized task behavior, where the adaptation steps needed for optimal performance in one task may not align with the requirements of other tasks. To address this, we propose a novel TTT approach called Synchronizing Tasks for Test-time Training (S4T), which enables the concurrent handling of multiple tasks. The core idea behind S4T is that predicting task relations across domain shifts is key to synchronizing tasks during test time. To validate our approach, we apply S4T to conventional multi-task benchmarks, integrating it with traditional TTT protocols. Our empirical results show that S4T outperforms state-of-the-art TTT methods across various benchmarks.
comment: Accepted at ICCV 2025
☆ A Unified Empirical Risk Minimization Framework for Flexible N-Tuples Weak Supervision
To alleviate the annotation burden in supervised learning, N-tuples learning has recently emerged as a powerful weakly-supervised method. While existing N-tuples learning approaches extend pairwise learning to higher-order comparisons and accommodate various real-world scenarios, they often rely on task-specific designs and lack a unified theoretical foundation. In this paper, we propose a general N-tuples learning framework based on empirical risk minimization, which systematically integrates pointwise unlabeled data to enhance learning performance. This paper first unifies the data generation processes of N-tuples and pointwise unlabeled data under a shared probabilistic formulation. Based on this unified view, we derive an unbiased empirical risk estimator that generalizes a broad class of existing N-tuples models. We further establish a generalization error bound for theoretical support. To demonstrate the flexibility of the framework, we instantiate it in four representative weakly supervised scenarios, each recoverable as a special case of our general model. Additionally, to address overfitting issues arising from negative risk terms, we adopt correction functions to adjust the empirical risk. Extensive experiments on benchmark datasets validate the effectiveness of the proposed framework and demonstrate that leveraging pointwise unlabeled data consistently improves generalization across various N-tuples learning tasks.
☆ BEAVER: Building Environments with Assessable Variation for Evaluating Multi-Objective Reinforcement Learning ICML
Recent years have seen significant advancements in designing reinforcement learning (RL)-based agents for building energy management. While individual success is observed in simulated or controlled environments, the scalability of RL approaches in terms of efficiency and generalization across building dynamics and operational scenarios remains an open question. In this work, we formally characterize the generalization space for the cross-environment, multi-objective building energy management task, and formulate the multi-objective contextual RL problem. Such a formulation helps understand the challenges of transferring learned policies across varied operational contexts such as climate and heat convection dynamics under multiple control objectives such as comfort level and energy consumption. We provide a principled way to parameterize such contextual information in realistic building RL environments, and construct a novel benchmark to facilitate the evaluation of generalizable RL algorithms in practical building control tasks. Our results show that existing multi-objective RL methods are capable of achieving reasonable trade-offs between conflicting objectives. However, their performance degrades under certain environment variations, underscoring the importance of incorporating dynamics-dependent contextual information into the policy learning process.
comment: Accepted at the Workshop on Computational Optimization of Buildings (ICML CO-BUILD), 42nd International Conference on Machine Learning (ICML 2025), Vancouver, Canada
☆ TRIX- Trading Adversarial Fairness via Mixed Adversarial Training
Adversarial Training (AT) is a widely adopted defense against adversarial examples. However, existing approaches typically apply a uniform training objective across all classes, overlooking disparities in class-wise vulnerability. This results in adversarial unfairness: classes with well distinguishable features (strong classes) tend to become more robust, while classes with overlapping or shared features(weak classes) remain disproportionately susceptible to adversarial attacks. We observe that strong classes do not require strong adversaries during training, as their non-robust features are quickly suppressed. In contrast, weak classes benefit from stronger adversaries to effectively reduce their vulnerabilities. Motivated by this, we introduce TRIX, a feature-aware adversarial training framework that adaptively assigns weaker targeted adversaries to strong classes, promoting feature diversity via uniformly sampled targets, and stronger untargeted adversaries to weak classes, enhancing their focused robustness. TRIX further incorporates per-class loss weighting and perturbation strength adjustments, building on prior work, to emphasize weak classes during the optimization. Comprehensive experiments on standard image classification benchmarks, including evaluations under strong attacks such as PGD and AutoAttack, demonstrate that TRIX significantly improves worst-case class accuracy on both clean and adversarial data, reducing inter-class robustness disparities, and preserves overall accuracy. Our results highlight TRIX as a practical step toward fair and effective adversarial defense.
☆ Distributed and Decentralised Training: Technical Governance Challenges in a Shifting AI Landscape ICML 2025
Advances in low-communication training algorithms are enabling a shift from centralised model training to compute setups that are either distributed across multiple clusters or decentralised via community-driven contributions. This paper distinguishes these two scenarios - distributed and decentralised training - which are little understood and often conflated in policy discourse. We discuss how they could impact technical AI governance through an increased risk of compute structuring, capability proliferation, and the erosion of detectability and shutdownability. While these trends foreshadow a possible new paradigm that could challenge key assumptions of compute governance, we emphasise that certain policy levers, like export controls, remain relevant. We also acknowledge potential benefits of decentralised AI, including privacy-preserving training runs that could unlock access to more data, and mitigating harmful power concentration. Our goal is to support more precise policymaking around compute, capability proliferation, and decentralised AI development.
comment: Accepted as an oral presentation at the Technical AI Governance Workshop (ICML 2025)
☆ OPC: One-Point-Contraction Unlearning Toward Deep Feature Forgetting
Machine unlearning seeks to remove the influence of particular data or class from trained models to meet privacy, legal, or ethical requirements. Existing unlearning methods tend to forget shallowly: phenomenon of an unlearned model pretend to forget by adjusting only the model response, while its internal representations retain information sufficiently to restore the forgotten data or behavior. We empirically confirm the widespread shallowness by reverting the forgetting effect of various unlearning methods via training-free performance recovery attack and gradient-inversion-based data reconstruction attack. To address this vulnerability fundamentally, we define a theoretical criterion of ``deep forgetting'' based on one-point-contraction of feature representations of data to forget. We also propose an efficient approximation algorithm, and use it to construct a novel general-purpose unlearning algorithm: One-Point-Contraction (OPC). Empirical evaluations on image classification unlearning benchmarks show that OPC achieves not only effective unlearning performance but also superior resilience against both performance recovery attack and gradient-inversion attack. The distinctive unlearning performance of OPC arises from the deep feature forgetting enforced by its theoretical foundation, and recaps the need for improved robustness of machine unlearning methods.
☆ Efficient and Scalable Estimation of Distributional Treatment Effects with Multi-Task Neural Networks
We propose a novel multi-task neural network approach for estimating distributional treatment effects (DTE) in randomized experiments. While DTE provides more granular insights into the experiment outcomes over conventional methods focusing on the Average Treatment Effect (ATE), estimating it with regression adjustment methods presents significant challenges. Specifically, precision in the distribution tails suffers due to data imbalance, and computational inefficiencies arise from the need to solve numerous regression problems, particularly in large-scale datasets commonly encountered in industry. To address these limitations, our method leverages multi-task neural networks to estimate conditional outcome distributions while incorporating monotonic shape constraints and multi-threshold label learning to enhance accuracy. To demonstrate the practical effectiveness of our proposed method, we apply our method to both simulated and real-world datasets, including a randomized field experiment aimed at reducing water consumption in the US and a large-scale A/B test from a leading streaming platform in Japan. The experimental results consistently demonstrate superior performance across various datasets, establishing our method as a robust and practical solution for modern causal inference applications requiring a detailed understanding of treatment effect heterogeneity.
☆ GuardVal: Dynamic Large Language Model Jailbreak Evaluation for Comprehensive Safety Testing
Jailbreak attacks reveal critical vulnerabilities in Large Language Models (LLMs) by causing them to generate harmful or unethical content. Evaluating these threats is particularly challenging due to the evolving nature of LLMs and the sophistication required in effectively probing their vulnerabilities. Current benchmarks and evaluation methods struggle to fully address these challenges, leaving gaps in the assessment of LLM vulnerabilities. In this paper, we review existing jailbreak evaluation practices and identify three assumed desiderata for an effective jailbreak evaluation protocol. To address these challenges, we introduce GuardVal, a new evaluation protocol that dynamically generates and refines jailbreak prompts based on the defender LLM's state, providing a more accurate assessment of defender LLMs' capacity to handle safety-critical situations. Moreover, we propose a new optimization method that prevents stagnation during prompt refinement, ensuring the generation of increasingly effective jailbreak prompts that expose deeper weaknesses in the defender LLMs. We apply this protocol to a diverse set of models, from Mistral-7b to GPT-4, across 10 safety domains. Our findings highlight distinct behavioral patterns among the models, offering a comprehensive view of their robustness. Furthermore, our evaluation process deepens the understanding of LLM behavior, leading to insights that can inform future research and drive the development of more secure models.
comment: 24 pages
☆ Adaptive Gaussian Mixture Models-based Anomaly Detection for under-constrained Cable-Driven Parallel Robots
Cable-Driven Parallel Robots (CDPRs) are increasingly used for load manipulation tasks involving predefined toolpaths with intermediate stops. At each stop, where the platform maintains a fixed pose and the motors keep the cables under tension, the system must evaluate whether it is safe to proceed by detecting anomalies that could compromise performance (e.g., wind gusts or cable impacts). This paper investigates whether anomalies can be detected using only motor torque data, without additional sensors. It introduces an adaptive, unsupervised outlier detection algorithm based on Gaussian Mixture Models (GMMs) to identify anomalies from torque signals. The method starts with a brief calibration period, just a few seconds, during which a GMM is fit on known anomaly-free data. Real-time torque measurements are then evaluated using Mahalanobis distance from the GMM, with statistically derived thresholds triggering anomaly flags. Model parameters are periodically updated using the latest segments identified as anomaly-free to adapt to changing conditions. Validation includes 14 long-duration test sessions simulating varied wind intensities. The proposed method achieves a 100% true positive rate and 95.4% average true negative rate, with 1-second detection latency. Comparative evaluation against power threshold and non-adaptive GMM methods indicates higher robustness to drift and environmental variation.
comment: 14 pages, 8 figures, 1 table, to be submitted to Advanced Intelligent Systems
☆ Balancing the Past and Present: A Coordinated Replay Framework for Federated Class-Incremental Learning
Federated Class Incremental Learning (FCIL) aims to collaboratively process continuously increasing incoming tasks across multiple clients. Among various approaches, data replay has become a promising solution, which can alleviate forgetting by reintroducing representative samples from previous tasks. However, their performance is typically limited by class imbalance, both within the replay buffer due to limited global awareness and between replayed and newly arrived classes. To address this issue, we propose a class wise balancing data replay method for FCIL (FedCBDR), which employs a global coordination mechanism for class-level memory construction and reweights the learning objective to alleviate the aforementioned imbalances. Specifically, FedCBDR has two key components: 1) the global-perspective data replay module reconstructs global representations of prior task in a privacy-preserving manner, which then guides a class-aware and importance-sensitive sampling strategy to achieve balanced replay; 2) Subsequently, to handle class imbalance across tasks, the task aware temperature scaling module adaptively adjusts the temperature of logits at both class and instance levels based on task dynamics, which reduces the model's overconfidence in majority classes while enhancing its sensitivity to minority classes. Experimental results verified that FedCBDR achieves balanced class-wise sampling under heterogeneous data distributions and improves generalization under task imbalance between earlier and recent tasks, yielding a 2%-15% Top-1 accuracy improvement over six state-of-the-art methods.
☆ Rationale-Enhanced Decoding for Multi-modal Chain-of-Thought
Large vision-language models (LVLMs) have demonstrated remarkable capabilities by integrating pre-trained vision encoders with large language models (LLMs). Similar to single-modal LLMs, chain-of-thought (CoT) prompting has been adapted for LVLMs to enhance multi-modal reasoning by generating intermediate rationales based on visual and textual inputs. While CoT is assumed to improve grounding and accuracy in LVLMs, our experiments reveal a key challenge: existing LVLMs often ignore the contents of generated rationales in CoT reasoning. To address this, we re-formulate multi-modal CoT reasoning as a KL-constrained reward maximization focused on rationale-conditional log-likelihood. As the optimal solution, we propose rationale-enhanced decoding (RED), a novel plug-and-play inference-time decoding strategy. RED harmonizes visual and rationale information by multiplying distinct image-conditional and rationale-conditional next token distributions. Extensive experiments show that RED consistently and significantly improves reasoning over standard CoT and other decoding methods across multiple benchmarks and LVLMs. Our work offers a practical and effective approach to improve both the faithfulness and accuracy of CoT reasoning in LVLMs, paving the way for more reliable rationale-grounded multi-modal systems.
comment: 17 pages, 4 figures
☆ Accelerating Transposed Convolutions on FPGA-based Edge Devices
Transposed Convolutions (TCONV) enable the up-scaling mechanism within generative Artificial Intelligence (AI) models. However, the predominant Input-Oriented Mapping (IOM) method for implementing TCONV has complex output mapping, overlapping sums, and ineffectual computations. These inefficiencies further exacerbate the performance bottleneck of TCONV and generative models on resource-constrained edge devices. To address this problem, in this paper we propose MM2IM, a hardware-software co-designed accelerator that combines Matrix Multiplication (MatMul) with col2IM to process TCONV layers on resource-constrained edge devices efficiently. Using the SECDA-TFLite design toolkit, we implement MM2IM and evaluate its performance across 261 TCONV problem configurations, achieving an average speedup of 1.9x against a dual-thread ARM Neon optimized CPU baseline. We then evaluate the performance of MM2IM on a range of TCONV layers from well-known generative models achieving up to 4.2x speedup, and compare it against similar resource-constrained TCONV accelerators, outperforming them by at least 2x GOPs/DSP. Finally, we evaluate MM2IM on the DCGAN and pix2pix GAN models, achieving up to 3x speedup and 2.4x energy reduction against the CPU baseline.
comment: Accepted to 35th International Conference on Field-Programmable Logic and Applications (FPL) 2025
☆ Some Theoretical Results on Layerwise Effective Dimension Oscillations in Finite Width ReLU Networks
We analyze the layerwise effective dimension (rank of the feature matrix) in fully-connected ReLU networks of finite width. Specifically, for a fixed batch of $m$ inputs and random Gaussian weights, we derive closed-form expressions for the expected rank of the \$m\times n\$ hidden activation matrices. Our main result shows that $\mathbb{E}[EDim(\ell)]=m[1-(1-2/\pi)^\ell]+O(e^{-c m})$ so that the rank deficit decays geometrically with ratio $1-2 / \pi \approx 0.3634$. We also prove a sub-Gaussian concentration bound, and identify the "revival" depths at which the expected rank attains local maxima. In particular, these peaks occur at depths $\ell_k^*\approx(k+1/2)\pi/\log(1/\rho)$ with height $\approx (1-e^{-\pi/2}) m \approx 0.79m$. We further show that this oscillatory rank behavior is a finite-width phenomenon: under orthogonal weight initialization or strong negative-slope leaky-ReLU, the rank remains (nearly) full. These results provide a precise characterization of how random ReLU layers alternately collapse and partially revive the subspace of input variations, adding nuance to prior work on expressivity of deep networks.
☆ Learning Pole Structures of Hadronic States using Predictive Uncertainty Estimation
Matching theoretical predictions to experimental data remains a central challenge in hadron spectroscopy. In particular, the identification of new hadronic states is difficult, as exotic signals near threshold can arise from a variety of physical mechanisms. A key diagnostic in this context is the pole structure of the scattering amplitude, but different configurations can produce similar signatures. The mapping between pole configurations and line shapes is especially ambiguous near the mass threshold, where analytic control is limited. In this work, we introduce an uncertainty-aware machine learning approach for classifying pole structures in $S$-matrix elements. Our method is based on an ensemble of classifier chains that provide both epistemic and aleatoric uncertainty estimates. We apply a rejection criterion based on predictive uncertainty, achieving a validation accuracy of nearly $95\%$ while discarding only a small fraction of high-uncertainty predictions. Trained on synthetic data with known pole structures, the model generalizes to previously unseen experimental data, including enhancements associated with the $P_{c\bar{c}}(4312)^+$ state observed by LHCb. In this, we infer a four-pole structure, representing the presence of a genuine compact pentaquark in the presence of a higher channel virtual state pole with non-vanishing width. While evaluated on this particular state, our framework is broadly applicable to other candidate hadronic states and offers a scalable tool for pole structure inference in scattering amplitudes.
☆ Machine Learning-Assisted Surrogate Modeling with Multi-Objective Optimization and Decision-Making of a Steam Methane Reforming Reactor
This study presents an integrated modeling and optimization framework for a steam methane reforming (SMR) reactor, combining a mathematical model, artificial neural network (ANN)-based hybrid modeling, advanced multi-objective optimization (MOO) and multi-criteria decision-making (MCDM) techniques. A one-dimensional fixed-bed reactor model accounting for internal mass transfer resistance was employed to simulate reactor performance. To reduce the high computational cost of the mathematical model, a hybrid ANN surrogate was constructed, achieving a 93.8% reduction in average simulation time while maintaining high predictive accuracy. The hybrid model was then embedded into three MOO scenarios using the non-dominated sorting genetic algorithm II (NSGA-II) solver: 1) maximizing methane conversion and hydrogen output; 2) maximizing hydrogen output while minimizing carbon dioxide emissions; and 3) a combined three-objective case. The optimal trade-off solutions were further ranked and selected using two MCDM methods: technique for order of preference by similarity to ideal solution (TOPSIS) and simplified preference ranking on the basis of ideal-average distance (sPROBID). Optimal results include a methane conversion of 0.863 with 4.556 mol/s hydrogen output in the first case, and 0.988 methane conversion with 3.335 mol/s hydrogen and 0.781 mol/s carbon dioxide in the third. This comprehensive methodology offers a scalable and effective strategy for optimizing complex catalytic reactor systems with multiple, often conflicting, objectives.
☆ HLF-FSL. A Decentralized Federated Split Learning Solution for IoT on Hyperledger Fabric
Collaborative machine learning in sensitive domains demands scalable, privacy preserving solutions for enterprise deployment. Conventional Federated Learning (FL) relies on a central server, introducing single points of failure and privacy risks, while Split Learning (SL) partitions models for privacy but scales poorly due to sequential training. We present a decentralized architecture that combines Federated Split Learning (FSL) with the permissioned blockchain Hyperledger Fabric (HLF). Our chaincode orchestrates FSL's split model execution and peer-to-peer aggregation without any central coordinator, leveraging HLF's transient fields and Private Data Collections (PDCs) to keep raw data and model activations private. On CIFAR-10 and MNIST benchmarks, HLF-FSL matches centralized FSL accuracy while reducing per epoch training time compared to Ethereum-based works. Performance and scalability tests show minimal blockchain overhead and preserved accuracy, demonstrating enterprise grade viability.
comment: 19 pages, 7 figures and 6 tables
☆ Exploring the Limits of Model Compression in LLMs: A Knowledge Distillation Study on QA Tasks
Large Language Models (LLMs) have demonstrated outstanding performance across a range of NLP tasks, however, their computational demands hinder their deployment in real-world, resource-constrained environments. This work investigates the extent to which LLMs can be compressed using Knowledge Distillation (KD) while maintaining strong performance on Question Answering (QA) tasks. We evaluate student models distilled from the Pythia and Qwen2.5 families on two QA benchmarks, SQuAD and MLQA, under zero-shot and one-shot prompting conditions. Results show that student models retain over 90% of their teacher models' performance while reducing parameter counts by up to 57.1%. Furthermore, one-shot prompting yields additional performance gains over zero-shot setups for both model families. These findings underscore the trade-off between model efficiency and task performance, demonstrating that KD, combined with minimal prompting, can yield compact yet capable QA systems suitable for resource-constrained applications.
comment: Accepted four publication at the 26th Meeting of the Special Interest on Discourse and Dialogue
☆ Concentration of measure for non-linear random matrices with applications to neural networks and non-commutative polynomials
We prove concentration inequalities for several models of non-linear random matrices. As corollaries we obtain estimates for linear spectral statistics of the conjugate kernel of neural networks and non-commutative polynomials in (possibly dependent) random matrices.
☆ TransformEEG: Towards Improving Model Generalizability in Deep Learning-based EEG Parkinson's Disease Detection
Electroencephalography (EEG) is establishing itself as an important, low-cost, noninvasive diagnostic tool for the early detection of Parkinson's Disease (PD). In this context, EEG-based Deep Learning (DL) models have shown promising results due to their ability to discover highly nonlinear patterns within the signal. However, current state-of-the-art DL models suffer from poor generalizability caused by high inter-subject variability. This high variability underscores the need for enhancing model generalizability by developing new architectures better tailored to EEG data. This paper introduces TransformEEG, a hybrid Convolutional-Transformer designed for Parkinson's disease detection using EEG data. Unlike transformer models based on the EEGNet structure, TransformEEG incorporates a depthwise convolutional tokenizer. This tokenizer is specialized in generating tokens composed by channel-specific features, which enables more effective feature mixing within the self-attention layers of the transformer encoder. To evaluate the proposed model, four public datasets comprising 290 subjects (140 PD patients, 150 healthy controls) were harmonized and aggregated. A 10-outer, 10-inner Nested-Leave-N-Subjects-Out (N-LNSO) cross-validation was performed to provide an unbiased comparison against seven other consolidated EEG deep learning models. TransformEEG achieved the highest balanced accuracy's median (78.45%) as well as the lowest interquartile range (6.37%) across all the N-LNSO partitions. When combined with data augmentation and threshold correction, median accuracy increased to 80.10%, with an interquartile range of 5.74%. In conclusion, TransformEEG produces more consistent and less skewed results. It demonstrates a substantial reduction in variability and more reliable PD detection using EEG data compared to the other investigated models.
comment: Submitted for possible publication. GitHub repository: see https://github.com/MedMaxLab/transformeeg
☆ Sparse Causal Discovery with Generative Intervention for Unsupervised Graph Domain Adaptation ICML 2025
Unsupervised Graph Domain Adaptation (UGDA) leverages labeled source domain graphs to achieve effective performance in unlabeled target domains despite distribution shifts. However, existing methods often yield suboptimal results due to the entanglement of causal-spurious features and the failure of global alignment strategies. We propose SLOGAN (Sparse Causal Discovery with Generative Intervention), a novel approach that achieves stable graph representation transfer through sparse causal modeling and dynamic intervention mechanisms. Specifically, SLOGAN first constructs a sparse causal graph structure, leveraging mutual information bottleneck constraints to disentangle sparse, stable causal features while compressing domain-dependent spurious correlations through variational inference. To address residual spurious correlations, we innovatively design a generative intervention mechanism that breaks local spurious couplings through cross-domain feature recombination while maintaining causal feature semantic consistency via covariance constraints. Furthermore, to mitigate error accumulation in target domain pseudo-labels, we introduce a category-adaptive dynamic calibration strategy, ensuring stable discriminative learning. Extensive experiments on multiple real-world datasets demonstrate that SLOGAN significantly outperforms existing baselines.
comment: ICML 2025
☆ Sparse Self-Federated Learning for Energy Efficient Cooperative Intelligence in Society 5.0
Federated Learning offers privacy-preserving collaborative intelligence but struggles to meet the sustainability demands of emerging IoT ecosystems necessary for Society 5.0-a human-centered technological future balancing social advancement with environmental responsibility. The excessive communication bandwidth and computational resources required by traditional FL approaches make them environmentally unsustainable at scale, creating a fundamental conflict with green AI principles as billions of resource-constrained devices attempt to participate. To this end, we introduce Sparse Proximity-based Self-Federated Learning (SParSeFuL), a resource-aware approach that bridges this gap by combining aggregate computing for self-organization with neural network sparsification to reduce energy and bandwidth consumption.
☆ Synthetic MC via Biological Transmitters: Therapeutic Modulation of the Gut-Brain Axis
Synthetic molecular communication (SMC) is a key enabler for future healthcare systems in which Internet of Bio-Nano-Things (IoBNT) devices facilitate the continuous monitoring of a patient's biochemical signals. To close the loop between sensing and actuation, both the detection and the generation of in-body molecular communication (MC) signals is key. However, generating signals inside the human body, e.g., via synthetic nanodevices, poses a challenge in SMC, due to technological obstacles as well as legal, safety, and ethical issues. Hence, this paper considers an SMC system in which signals are generated indirectly via the modulation of a natural in-body MC system, namely the gut-brain axis (GBA). Therapeutic GBA modulation is already established as treatment for neurological diseases, e.g., drug refractory epilepsy (DRE), and performed via the administration of nutritional supplements or specific diets. However, the molecular signaling pathways that mediate the effect of such treatments are mostly unknown. Consequently, existing treatments are standardized or designed heuristically and able to help only some patients while failing to help others. In this paper, we propose to leverage personal health data, e.g., gathered by in-body IoBNT devices, to design more versatile and robust GBA modulation-based treatments as compared to the existing ones. To show the feasibility of our approach, we define a catalog of theoretical requirements for therapeutic GBA modulation. Then, we propose a machine learning model to verify these requirements for practical scenarios when only limited data on the GBA modulation exists. By evaluating the proposed model on several datasets, we confirm its excellent accuracy in identifying different modulators of the GBA. Finally, we utilize the proposed model to identify specific modulatory pathways that play an important role for therapeutic GBA modulation.
☆ Stress Monitoring in Healthcare: An Ensemble Machine Learning Framework Using Wearable Sensor Data
Healthcare professionals, particularly nurses, face elevated occupational stress, a concern amplified during the COVID-19 pandemic. While wearable sensors offer promising avenues for real-time stress monitoring, existing studies often lack comprehensive datasets and robust analytical frameworks. This study addresses these gaps by introducing a multimodal dataset comprising physiological signals, electrodermal activity, heart rate and skin temperature. A systematic literature review identified limitations in prior stress-detection methodologies, particularly in handling class imbalance and optimizing model generalizability. To overcome these challenges, the dataset underwent preprocessing with the Synthetic Minority Over sampling Technique (SMOTE), ensuring balanced representation of stress states. Advanced machine learning models including Random Forest, XGBoost and a Multi-Layer Perceptron (MLP) were evaluated and combined into a Stacking Classifier to leverage their collective predictive strengths. By using a publicly accessible dataset and a reproducible analytical pipeline, this work advances the development of deployable stress-monitoring systems, offering practical implications for safeguarding healthcare workers' mental health. Future research directions include expanding demographic diversity and exploring edge-computing implementations for low latency stress alerts.
☆ Bayesian Discrete Diffusion Beats Autoregressive Perplexity
We reveal a hidden Bayesian core of discrete-diffusion language models by showing that the expected denoiser output under the forward masking distribution recovers the exact posterior over clean tokens. Under minimal assumptions, Monte Carlo marginalization over K independent corruptions converges to this posterior at rate O(1/sqrt(K)), yielding a simple proof of consistency and finite-sample error bounds. Building on this insight, we introduce a lightweight inference-time ensemble that averages K mask-and-denoise passes to obtain posterior-aware token probabilities and uncertainty estimates at no extra training cost. On WikiText-2, our method achieves test perplexity 8.8 with K=8, versus 20.3 for GPT-2 Small, despite using a model of comparable size. Code is available at https://github.com/mercury0100/bayesradd.
comment: 12 pages, 2 figures, 2 tables
☆ Improving Clustering on Occupational Text Data through Dimensionality Reduction
In this study, we focused on proposing an optimal clustering mechanism for the occupations defined in the well-known US-based occupational database, O*NET. Even though all occupations are defined according to well-conducted surveys in the US, their definitions can vary for different firms and countries. Hence, if one wants to expand the data that is already collected in O*NET for the occupations defined with different tasks, a map between the definitions will be a vital requirement. We proposed a pipeline using several BERT-based techniques with various clustering approaches to obtain such a map. We also examined the effect of dimensionality reduction approaches on several metrics used in measuring performance of clustering algorithms. Finally, we improved our results by using a specialized silhouette approach. This new clustering-based mapping approach with dimensionality reduction may help distinguish the occupations automatically, creating new paths for people wanting to change their careers.
comment: Preprint, 10 figures
☆ CHOMET: Conditional Handovers via Meta-Learning
Handovers (HOs) are the cornerstone of modern cellular networks for enabling seamless connectivity to a vast and diverse number of mobile users. However, as mobile networks become more complex with more diverse users and smaller cells, traditional HOs face significant challenges, such as prolonged delays and increased failures. To mitigate these issues, 3GPP introduced conditional handovers (CHOs), a new type of HO that enables the preparation (i.e., resource allocation) of multiple cells for a single user to increase the chance of HO success and decrease the delays in the procedure. Despite its advantages, CHO introduces new challenges that must be addressed, including efficient resource allocation and managing signaling/communication overhead from frequent cell preparations and releases. This paper presents a novel framework aligned with the O-RAN paradigm that leverages meta-learning for CHO optimization, providing robust dynamic regret guarantees and demonstrating at least 180% superior performance than other 3GPP benchmarks in volatile signal conditions.
☆ COALA: Numerically Stable and Efficient Framework for Context-Aware Low-Rank Approximation
Recent studies suggest that context-aware low-rank approximation is a useful tool for compression and fine-tuning of modern large-scale neural networks. In this type of approximation, a norm is weighted by a matrix of input activations, significantly improving metrics over the unweighted case. Nevertheless, existing methods for neural networks suffer from numerical instabilities due to their reliance on classical formulas involving explicit Gram matrix computation and their subsequent inversion. We demonstrate that this can degrade the approximation quality or cause numerically singular matrices. To address these limitations, we propose a novel inversion-free regularized framework that is based entirely on stable decompositions and overcomes the numerical pitfalls of prior art. Our method can handle possible challenging scenarios: (1) when calibration matrices exceed GPU memory capacity, (2) when input activation matrices are nearly singular, and even (3) when insufficient data prevents unique approximation. For the latter, we prove that our solution converges to a desired approximation and derive explicit error bounds.
☆ On Trustworthy Rule-Based Models and Explanations
A task of interest in machine learning (ML) is that of ascribing explanations to the predictions made by ML models. Furthermore, in domains deemed high risk, the rigor of explanations is paramount. Indeed, incorrect explanations can and will mislead human decision makers. As a result, and even if interpretability is acknowledged as an elusive concept, so-called interpretable models are employed ubiquitously in high-risk uses of ML and data mining (DM). This is the case for rule-based ML models, which encompass decision trees, diagrams, sets and lists. This paper relates explanations with well-known undesired facets of rule-based ML models, which include negative overlap and several forms of redundancy. The paper develops algorithms for the analysis of these undesired facets of rule-based systems, and concludes that well-known and widely used tools for learning rule-based ML models will induce rule sets that exhibit one or more negative facets.
☆ Real-Time Decorrelation-Based Anomaly Detection for Multivariate Time Series
Anomaly detection (AD) plays a vital role across a wide range of real-world domains by identifying data instances that deviate from expected patterns, potentially signaling critical events such as system failures, fraudulent activities, or rare medical conditions. The demand for real-time AD has surged with the rise of the (Industrial) Internet of Things, where massive volumes of multivariate sensor data must be processed instantaneously. Real-time AD requires methods that not only handle high-dimensional streaming data but also operate in a single-pass manner, without the burden of storing historical instances, thereby ensuring minimal memory usage and fast decision-making. We propose DAD, a novel real-time decorrelation-based anomaly detection method for multivariate time series, based on an online decorrelation learning approach. Unlike traditional proximity-based or reconstruction-based detectors that process entire data or windowed instances, DAD dynamically learns and monitors the correlation structure of data sample by sample in a single pass, enabling efficient and effective detection. To support more realistic benchmarking practices, we also introduce a practical hyperparameter tuning strategy tailored for real-time anomaly detection scenarios. Extensive experiments on widely used benchmark datasets demonstrate that DAD achieves the most consistent and superior performance across diverse anomaly types compared to state-of-the-art methods. Crucially, its robustness to increasing dimensionality makes it particularly well-suited for real-time, high-dimensional data streams. Ultimately, DAD not only strikes an optimal balance between detection efficacy and computational efficiency but also sets a new standard for real-time, memory-constrained anomaly detection.
☆ Position: We Need An Algorithmic Understanding of Generative AI ICML 2025
What algorithms do LLMs actually learn and use to solve problems? Studies addressing this question are sparse, as research priorities are focused on improving performance through scale, leaving a theoretical and empirical gap in understanding emergent algorithms. This position paper proposes AlgEval: a framework for systematic research into the algorithms that LLMs learn and use. AlgEval aims to uncover algorithmic primitives, reflected in latent representations, attention, and inference-time compute, and their algorithmic composition to solve task-specific problems. We highlight potential methodological paths and a case study toward this goal, focusing on emergent search algorithms. Our case study illustrates both the formation of top-down hypotheses about candidate algorithms, and bottom-up tests of these hypotheses via circuit-level analysis of attention patterns and hidden states. The rigorous, systematic evaluation of how LLMs actually solve tasks provides an alternative to resource-intensive scaling, reorienting the field toward a principled understanding of underlying computations. Such algorithmic explanations offer a pathway to human-understandable interpretability, enabling comprehension of the model's internal reasoning performance measures. This can in turn lead to more sample-efficient methods for training and improving performance, as well as novel architectures for end-to-end and multi-agent systems.
comment: Accepted at ICML 2025 as a Spotlight Position Paper
☆ Neural Concept Verifier: Scaling Prover-Verifier Games via Concept Encodings
While Prover-Verifier Games (PVGs) offer a promising path toward verifiability in nonlinear classification models, they have not yet been applied to complex inputs such as high-dimensional images. Conversely, Concept Bottleneck Models (CBMs) effectively translate such data into interpretable concepts but are limited by their reliance on low-capacity linear predictors. In this work, we introduce the Neural Concept Verifier (NCV), a unified framework combining PVGs with concept encodings for interpretable, nonlinear classification in high-dimensional settings. NCV achieves this by utilizing recent minimally supervised concept discovery models to extract structured concept encodings from raw inputs. A prover then selects a subset of these encodings, which a verifier -- implemented as a nonlinear predictor -- uses exclusively for decision-making. Our evaluations show that NCV outperforms CBM and pixel-based PVG classifier baselines on high-dimensional, logically complex datasets and also helps mitigate shortcut behavior. Overall, we demonstrate NCV as a promising step toward performative, verifiable AI.
comment: 16 pages, 4 figures, 8 tables
☆ Uncertainty Quantification for Motor Imagery BCI -- Machine Learning vs. Deep Learning
Brain-computer interfaces (BCIs) turn brain signals into functionally useful output, but they are not always accurate. A good Machine Learning classifier should be able to indicate how confident it is about a given classification, by giving a probability for its classification. Standard classifiers for Motor Imagery BCIs do give such probabilities, but research on uncertainty quantification has been limited to Deep Learning. We compare the uncertainty quantification ability of established BCI classifiers using Common Spatial Patterns (CSP-LDA) and Riemannian Geometry (MDRM) to specialized methods in Deep Learning (Deep Ensembles and Direct Uncertainty Quantification) as well as standard Convolutional Neural Networks (CNNs). We found that the overconfidence typically seen in Deep Learning is not a problem in CSP-LDA and MDRM. We found that MDRM is underconfident, which we solved by adding Temperature Scaling (MDRM-T). CSP-LDA and MDRM-T give the best uncertainty estimates, but Deep Ensembles and standard CNNs give the best classifications. We show that all models are able to separate between easy and difficult estimates, so that we can increase the accuracy of a Motor Imagery BCI by rejecting samples that are ambiguous.
comment: 6 pages, 3 figures
☆ Divergence Minimization Preference Optimization for Diffusion Model Alignment
Diffusion models have achieved remarkable success in generating realistic and versatile images from text prompts. Inspired by the recent advancements of language models, there is an increasing interest in further improving the models by aligning with human preferences. However, we investigate alignment from a divergence minimization perspective and reveal that existing preference optimization methods are typically trapped in suboptimal mean-seeking optimization. In this paper, we introduce Divergence Minimization Preference Optimization (DMPO), a novel and principled method for aligning diffusion models by minimizing reverse KL divergence, which asymptotically enjoys the same optimization direction as original RL. We provide rigorous analysis to justify the effectiveness of DMPO and conduct comprehensive experiments to validate its empirical strength across both human evaluations and automatic metrics. Our extensive results show that diffusion models fine-tuned with DMPO can consistently outperform or match existing techniques, specifically outperforming all existing diffusion alignment baselines by at least 64.6% in PickScore across all evaluation datasets, demonstrating the method's superiority in aligning generative behavior with desired outputs. Overall, DMPO unlocks a robust and elegant pathway for preference alignment, bridging principled theory with practical performance in diffusion models.
comment: 24 pages, 8 figures
☆ Teaching LLM to Reason: Reinforcement Learning from Algorithmic Problems without Code
Enhancing reasoning capabilities remains a central focus in the LLM reasearch community. A promising direction involves requiring models to simulate code execution step-by-step to derive outputs for given inputs. However, as code is often designed for large-scale systems, direct application leads to over-reliance on complex data structures and algorithms, even for simple cases, resulting in overfitting to algorithmic patterns rather than core reasoning structures. To address this, we propose TeaR, which aims at teaching LLMs to reason better. TeaR leverages careful data curation and reinforcement learning to guide models in discovering optimal reasoning paths through code-related tasks, thereby improving general reasoning abilities. We conduct extensive experiments using two base models and three long-CoT distillation models, with model sizes ranging from 1.5 billion to 32 billion parameters, and across 17 benchmarks spanning Math, Knowledge, Code, and Logical Reasoning. The results consistently show significant performance improvements. Notably, TeaR achieves a 35.9% improvement on Qwen2.5-7B and 5.9% on R1-Distilled-7B.
☆ Semi-supervised learning and integration of multi-sequence MR-images for carotid vessel wall and plaque segmentation
The analysis of carotid arteries, particularly plaques, in multi-sequence Magnetic Resonance Imaging (MRI) data is crucial for assessing the risk of atherosclerosis and ischemic stroke. In order to evaluate metrics and radiomic features, quantifying the state of atherosclerosis, accurate segmentation is important. However, the complex morphology of plaques and the scarcity of labeled data poses significant challenges. In this work, we address these problems and propose a semi-supervised deep learning-based approach designed to effectively integrate multi-sequence MRI data for the segmentation of carotid artery vessel wall and plaque. The proposed algorithm consists of two networks: a coarse localization model identifies the region of interest guided by some prior knowledge on the position and number of carotid arteries, followed by a fine segmentation model for precise delineation of vessel walls and plaques. To effectively integrate complementary information across different MRI sequences, we investigate different fusion strategies and introduce a multi-level multi-sequence version of U-Net architecture. To address the challenges of limited labeled data and the complexity of carotid artery MRI, we propose a semi-supervised approach that enforces consistency under various input transformations. Our approach is evaluated on 52 patients with arteriosclerosis, each with five MRI sequences. Comprehensive experiments demonstrate the effectiveness of our approach and emphasize the role of fusion point selection in U-Net-based architectures. To validate the accuracy of our results, we also include an expert-based assessment of model performance. Our findings highlight the potential of fusion strategies and semi-supervised learning for improving carotid artery segmentation in data-limited MRI applications.
☆ Resolving Token-Space Gradient Conflicts: Token Space Manipulation for Transformer-Based Multi-Task Learning ICCV 2025
Multi-Task Learning (MTL) enables multiple tasks to be learned within a shared network, but differences in objectives across tasks can cause negative transfer, where the learning of one task degrades another task's performance. While pre-trained transformers significantly improve MTL performance, their fixed network capacity and rigid structure limit adaptability. Previous dynamic network architectures attempt to address this but are inefficient as they directly convert shared parameters into task-specific ones. We propose Dynamic Token Modulation and Expansion (DTME-MTL), a framework applicable to any transformer-based MTL architecture. DTME-MTL enhances adaptability and reduces overfitting by identifying gradient conflicts in token space and applying adaptive solutions based on conflict type. Unlike prior methods that mitigate negative transfer by duplicating network parameters, DTME-MTL operates entirely in token space, enabling efficient adaptation without excessive parameter growth. Extensive experiments demonstrate that DTME-MTL consistently improves multi-task performance with minimal computational overhead, offering a scalable and effective solution for enhancing transformer-based MTL models.
comment: Accepted at ICCV 2025
☆ Machine Bullshit: Characterizing the Emergent Disregard for Truth in Large Language Models
Bullshit, as conceptualized by philosopher Harry Frankfurt, refers to statements made without regard to their truth value. While previous work has explored large language model (LLM) hallucination and sycophancy, we propose machine bullshit as an overarching conceptual framework that can allow researchers to characterize the broader phenomenon of emergent loss of truthfulness in LLMs and shed light on its underlying mechanisms. We introduce the Bullshit Index, a novel metric quantifying LLMs' indifference to truth, and propose a complementary taxonomy analyzing four qualitative forms of bullshit: empty rhetoric, paltering, weasel words, and unverified claims. We conduct empirical evaluations on the Marketplace dataset, the Political Neutrality dataset, and our new BullshitEval benchmark (2,400 scenarios spanning 100 AI assistants) explicitly designed to evaluate machine bullshit. Our results demonstrate that model fine-tuning with reinforcement learning from human feedback (RLHF) significantly exacerbates bullshit and inference-time chain-of-thought (CoT) prompting notably amplify specific bullshit forms, particularly empty rhetoric and paltering. We also observe prevalent machine bullshit in political contexts, with weasel words as the dominant strategy. Our findings highlight systematic challenges in AI alignment and provide new insights toward more truthful LLM behavior.
comment: Project page, code & data: https://machine-bullshit.github.io
☆ Galerkin-ARIMA: A Two-Stage Polynomial Regression Framework for Fast Rolling One-Step-Ahead Forecasting
Time-series models like ARIMA remain widely used for forecasting but limited to linear assumptions and high computational cost in large and complex datasets. We propose Galerkin-ARIMA that generalizes the AR component of ARIMA and replace it with a flexible spline-based function estimated by Galerkin projection. This enables the model to capture nonlinear dependencies in lagged values and retain the MA component and Gaussian noise assumption. We derive a closed-form OLS estimator for the Galerkin coefficients and show the model is asymptotically unbiased and consistent under standard conditions. Our method bridges classical time-series modeling and nonparametric regression, which offering improved forecasting performance and computational efficiency.
☆ Hess-MC2: Sequential Monte Carlo Squared using Hessian Information and Second Order Proposals
When performing Bayesian inference using Sequential Monte Carlo (SMC) methods, two considerations arise: the accuracy of the posterior approximation and computational efficiency. To address computational demands, Sequential Monte Carlo Squared (SMC$^2$) is well-suited for high-performance computing (HPC) environments. The design of the proposal distribution within SMC$^2$ can improve accuracy and exploration of the posterior as poor proposals may lead to high variance in importance weights and particle degeneracy. The Metropolis-Adjusted Langevin Algorithm (MALA) uses gradient information so that particles preferentially explore regions of higher probability. In this paper, we extend this idea by incorporating second-order information, specifically the Hessian of the log-target. While second-order proposals have been explored previously in particle Markov Chain Monte Carlo (p-MCMC) methods, we are the first to introduce them within the SMC$^2$ framework. Second-order proposals not only use the gradient (first-order derivative), but also the curvature (second-order derivative) of the target distribution. Experimental results on synthetic models highlight the benefits of our approach in terms of step-size selection and posterior approximation accuracy when compared to other proposals.
comment: Accepted to IEEE Machine Learning Signal Processing conference 2025
☆ General purpose models for the chemical sciences
Data-driven techniques have a large potential to transform and accelerate the chemical sciences. However, chemical sciences also pose the unique challenge of very diverse, small, fuzzy datasets that are difficult to leverage in conventional machine learning approaches completely. A new class of models, general-purpose models (GPMs) such as large language models, have shown the ability to solve tasks they have not been directly trained on, and to flexibly operate with low amounts of data in different formats. In this review, we discuss fundamental building principles of GPMs and review recent applications of those models in the chemical sciences across the entire scientific process. While many of these applications are still in the prototype phase, we expect that the increasing interest in GPMs will make many of them mature in the coming years.
☆ Neural networks leverage nominally quantum and post-quantum representations
We show that deep neural networks, including transformers and RNNs, pretrained as usual on next-token prediction, intrinsically discover and represent beliefs over 'quantum' and 'post-quantum' low-dimensional generative models of their training data -- as if performing iterative Bayesian updates over the latent state of this world model during inference as they observe more context. Notably, neural nets easily find these representation whereas there is no finite classical circuit that would do the job. The corresponding geometric relationships among neural activations induced by different input sequences are found to be largely independent of neural-network architecture. Each point in this geometry corresponds to a history-induced probability density over all possible futures, and the relative displacement of these points reflects the difference in mechanism and magnitude for how these distinct pasts affect the future.
☆ Probabilistic Approximate Optimization: A New Variational Monte Carlo Algorithm
We introduce a generalized \textit{Probabilistic Approximate Optimization Algorithm (PAOA)}, a classical variational Monte Carlo framework that extends and formalizes prior work by Weitz \textit{et al.}~\cite{Combes_2023}, enabling parameterized and fast sampling on present-day Ising machines and probabilistic computers. PAOA operates by iteratively modifying the couplings of a network of binary stochastic units, guided by cost evaluations from independent samples. We establish a direct correspondence between derivative-free updates and the gradient of the full $2^N \times 2^N$ Markov flow, showing that PAOA admits a principled variational formulation. Simulated annealing emerges as a limiting case under constrained parameterizations, and we implement this regime on an FPGA-based probabilistic computer with on-chip annealing to solve large 3D spin-glass problems. Benchmarking PAOA against QAOA on the canonical 26-spin Sherrington-Kirkpatrick model with matched parameters reveals superior performance for PAOA. We show that PAOA naturally extends simulated annealing by optimizing multiple temperature profiles, leading to improved performance over SA on heavy-tailed problems such as SK-L\'evy.
☆ Autonomous AI-based Cybersecurity Framework for Critical Infrastructure: Real-Time Threat Mitigation
Critical infrastructure systems, including energy grids, healthcare facilities, transportation networks, and water distribution systems, are pivotal to societal stability and economic resilience. However, the increasing interconnectivity of these systems exposes them to various cyber threats, including ransomware, Denial-of-Service (DoS) attacks, and Advanced Persistent Threats (APTs). This paper examines cybersecurity vulnerabilities in critical infrastructure, highlighting the threat landscape, attack vectors, and the role of Artificial Intelligence (AI) in mitigating these risks. We propose a hybrid AI-driven cybersecurity framework to enhance real-time vulnerability detection, threat modelling, and automated remediation. This study also addresses the complexities of adversarial AI, regulatory compliance, and integration. Our findings provide actionable insights to strengthen the security and resilience of critical infrastructure systems against emerging cyber threats.
comment: 7 pages, IEEE conference
☆ Hybrid LLM-Enhanced Intrusion Detection for Zero-Day Threats in IoT Networks
This paper presents a novel approach to intrusion detection by integrating traditional signature-based methods with the contextual understanding capabilities of the GPT-2 Large Language Model (LLM). As cyber threats become increasingly sophisticated, particularly in distributed, heterogeneous, and resource-constrained environments such as those enabled by the Internet of Things (IoT), the need for dynamic and adaptive Intrusion Detection Systems (IDSs) becomes increasingly urgent. While traditional methods remain effective for detecting known threats, they often fail to recognize new and evolving attack patterns. In contrast, GPT-2 excels at processing unstructured data and identifying complex semantic relationships, making it well-suited to uncovering subtle, zero-day attack vectors. We propose a hybrid IDS framework that merges the robustness of signature-based techniques with the adaptability of GPT-2-driven semantic analysis. Experimental evaluations on a representative intrusion dataset demonstrate that our model enhances detection accuracy by 6.3%, reduces false positives by 9.0%, and maintains near real-time responsiveness. These results affirm the potential of language model integration to build intelligent, scalable, and resilient cybersecurity defences suited for modern connected environments.
comment: 6 pages, IEEE conference
☆ Phishing Detection in the Gen-AI Era: Quantized LLMs vs Classical Models
Phishing attacks are becoming increasingly sophisticated, underscoring the need for detection systems that strike a balance between high accuracy and computational efficiency. This paper presents a comparative evaluation of traditional Machine Learning (ML), Deep Learning (DL), and quantized small-parameter Large Language Models (LLMs) for phishing detection. Through experiments on a curated dataset, we show that while LLMs currently underperform compared to ML and DL methods in terms of raw accuracy, they exhibit strong potential for identifying subtle, context-based phishing cues. We also investigate the impact of zero-shot and few-shot prompting strategies, revealing that LLM-rephrased emails can significantly degrade the performance of both ML and LLM-based detectors. Our benchmarking highlights that models like DeepSeek R1 Distill Qwen 14B (Q8_0) achieve competitive accuracy, above 80%, using only 17GB of VRAM, supporting their viability for cost-efficient deployment. We further assess the models' adversarial robustness and cost-performance tradeoffs, and demonstrate how lightweight LLMs can provide concise, interpretable explanations to support real-time decision-making. These findings position optimized LLMs as promising components in phishing defence systems and offer a path forward for integrating explainable, efficient AI into modern cybersecurity frameworks.
comment: 8 Pages, IEEE Conference
☆ HGMP:Heterogeneous Graph Multi-Task Prompt Learning IJCAI-25
The pre-training and fine-tuning methods have gained widespread attention in the field of heterogeneous graph neural networks due to their ability to leverage large amounts of unlabeled data during the pre-training phase, allowing the model to learn rich structural features. However, these methods face the issue of a mismatch between the pre-trained model and downstream tasks, leading to suboptimal performance in certain application scenarios. Prompt learning methods have emerged as a new direction in heterogeneous graph tasks, as they allow flexible adaptation of task representations to address target inconsistency. Building on this idea, this paper proposes a novel multi-task prompt framework for the heterogeneous graph domain, named HGMP. First, to bridge the gap between the pre-trained model and downstream tasks, we reformulate all downstream tasks into a unified graph-level task format. Next, we address the limitations of existing graph prompt learning methods, which struggle to integrate contrastive pre-training strategies in the heterogeneous graph domain. We design a graph-level contrastive pre-training strategy to better leverage heterogeneous information and enhance performance in multi-task scenarios. Finally, we introduce heterogeneous feature prompts, which enhance model performance by refining the representation of input graph features. Experimental results on public datasets show that our proposed method adapts well to various tasks and significantly outperforms baseline methods.
comment: The 25th International Joint Conference on Artificial Intelligence (IJCAI-25)
☆ Generalized Tree Edit Distance (GTED): A Faithful Evaluation Metric for Statement Autoformalization ICML25
Statement autoformalization, the automated translation of statement from natural language into formal languages, has become a subject of extensive research, yet the development of robust automated evaluation metrics remains limited. Existing evaluation methods often lack semantic understanding, face challenges with high computational costs, and are constrained by the current progress of automated theorem proving. To address these issues, we propose GTED (Generalized Tree Edit Distance), a novel evaluation framework that first standardizes formal statements and converts them into operator trees, then determines the semantic similarity using the eponymous GTED metric. On the miniF2F and ProofNet benchmarks, GTED outperforms all baseline metrics by achieving the highest accuracy and Kappa scores, thus providing the community with a more faithful metric for automated evaluation. The code and experimental results are available at https://github.com/XiaoyangLiu-sjtu/GTED.
comment: Accepted to AI4Math@ICML25
☆ IML-Spikeformer: Input-aware Multi-Level Spiking Transformer for Speech Processing
Spiking Neural Networks (SNNs), inspired by biological neural mechanisms, represent a promising neuromorphic computing paradigm that offers energy-efficient alternatives to traditional Artificial Neural Networks (ANNs). Despite proven effectiveness, SNN architectures have struggled to achieve competitive performance on large-scale speech processing task. Two key challenges hinder progress: (1) the high computational overhead during training caused by multi-timestep spike firing, and (2) the absence of large-scale SNN architectures tailored to speech processing tasks. To overcome the issues, we introduce Input-aware Multi-Level Spikeformer, i.e. IML-Spikeformer, a spiking Transformer architecture specifically designed for large-scale speech processing. Central to our design is the Input-aware Multi-Level Spike (IMLS) mechanism, which simulate multi-timestep spike firing within a single timestep using an adaptive, input-aware thresholding scheme. IML-Spikeformer further integrates a Reparameterized Spiking Self-Attention (RepSSA) module with a Hierarchical Decay Mask (HDM), forming the HD-RepSSA module. This module enhances the precision of attention maps and enables modeling of multi-scale temporal dependencies in speech signals. Experiments demonstrate that IML-Spikeformer achieves word error rates of 6.0\% on AiShell-1 and 3.4\% on Librispeech-960, comparable to conventional ANN transformers while reducing theoretical inference energy consumption by 4.64$\times$ and 4.32$\times$ respectively. IML-Spikeformer marks an advance of scalable SNN architectures for large-scale speech processing in both task performance and energy efficiency.
comment: Under review of TNNLS
☆ Learning Collective Variables from Time-lagged Generation
Rare events such as state transitions are difficult to observe directly with molecular dynamics simulations due to long timescales. Enhanced sampling techniques overcome this by introducing biases along carefully chosen low-dimensional features, known as collective variables (CVs), which capture the slow degrees of freedom. Machine learning approaches (MLCVs) have automated CV discovery, but existing methods typically focus on discriminating meta-stable states without fully encoding the detailed dynamics essential for accurate sampling. We propose TLC, a framework that learns CVs directly from time-lagged conditions of a generative model. Instead of modeling the static Boltzmann distribution, TLC models a time-lagged conditional distribution yielding CVs to capture the slow dynamic behavior. We validate TLC on the Alanine Dipeptide system using two CV-based enhanced sampling tasks: (i) steered molecular dynamics (SMD) and (ii) on-the-fly probability enhanced sampling (OPES), demonstrating equal or superior performance compared to existing MLCV methods in both transition path sampling and state discrimination.
☆ ST-GRIT: Spatio-Temporal Graph Transformer For Internal Ice Layer Thickness Prediction ICIP
Understanding the thickness and variability of internal ice layers in radar imagery is crucial for monitoring snow accumulation, assessing ice dynamics, and reducing uncertainties in climate models. Radar sensors, capable of penetrating ice, provide detailed radargram images of these internal layers. In this work, we present ST-GRIT, a spatio-temporal graph transformer for ice layer thickness, designed to process these radargrams and capture the spatiotemporal relationships between shallow and deep ice layers. ST-GRIT leverages an inductive geometric graph learning framework to extract local spatial features as feature embeddings and employs a series of temporal and spatial attention blocks separately to model long-range dependencies effectively in both dimensions. Experimental evaluation on radargram data from the Greenland ice sheet demonstrates that ST-GRIT consistently outperforms current state-of-the-art methods and other baseline graph neural networks by achieving lower root mean-squared error. These results highlight the advantages of self-attention mechanisms on graphs over pure graph neural networks, including the ability to handle noise, avoid oversmoothing, and capture long-range dependencies. Moreover, the use of separate spatial and temporal attention blocks allows for distinct and robust learning of spatial relationships and temporal patterns, providing a more comprehensive and effective approach.
comment: Accepted for 2025 IEEE International Conference on Image Processing (ICIP)
☆ GRIT: Graph Transformer For Internal Ice Layer Thickness Prediction
Gaining a deeper understanding of the thickness and variability of internal ice layers in Radar imagery is essential in monitoring the snow accumulation, better evaluating ice dynamics processes, and minimizing uncertainties in climate models. Radar sensors, capable of penetrating ice, capture detailed radargram images of internal ice layers. In this work, we introduce GRIT, graph transformer for ice layer thickness. GRIT integrates an inductive geometric graph learning framework with an attention mechanism, designed to map the relationships between shallow and deeper ice layers. Compared to baseline graph neural networks, GRIT demonstrates consistently lower prediction errors. These results highlight the attention mechanism's effectiveness in capturing temporal changes across ice layers, while the graph transformer combines the strengths of transformers for learning long-range dependencies with graph neural networks for capturing spatial patterns, enabling robust modeling of complex spatiotemporal dynamics.
comment: Accepted for 2025 IEEE International Geoscience and Remote Sensing Symposium (IGARSS 2025)
☆ Simple Mechanistic Explanations for Out-Of-Context Reasoning ICML 2025
Out-of-context reasoning (OOCR) is a phenomenon in which fine-tuned LLMs exhibit surprisingly deep out-of-distribution generalization. Rather than learning shallow heuristics, they implicitly internalize and act on the consequences of observations scattered throughout the fine-tuning data. In this work, we investigate this phenomenon mechanistically and find that many instances of OOCR in the literature have a simple explanation: the LoRA fine-tuning essentially adds a constant steering vector, steering the model towards a general concept. This improves performance on the fine-tuning task and in many other concept-related domains, causing the surprising generalization. Moreover, we can directly train steering vectors for these tasks from scratch, which also induces OOCR. We find that our results hold even for a task that seems like it must involve conditional behavior (model backdoors); it turns out that unconditionally adding a steering vector is sufficient. Overall, our work presents one explanation of what gets learned during fine-tuning for OOCR tasks, contributing to the key question of why LLMs can reason out of context, an advanced capability that is highly relevant to their safe and reliable deployment.
comment: ICML 2025 Workshop R2-FM
♻ ☆ KinDEL: DNA-Encoded Library Dataset for Kinase Inhibitors
DNA-Encoded Libraries (DELs) represent a transformative technology in drug discovery, facilitating the high-throughput exploration of vast chemical spaces. Despite their potential, the scarcity of publicly available DEL datasets presents a bottleneck for the advancement of machine learning methodologies in this domain. To address this gap, we introduce KinDEL, one of the largest publicly accessible DEL datasets and the first one that includes binding poses from molecular docking experiments. Focused on two kinases, Mitogen-Activated Protein Kinase 14 (MAPK14) and Discoidin Domain Receptor Tyrosine Kinase 1 (DDR1), KinDEL includes 81 million compounds, offering a rich resource for computational exploration. Additionally, we provide comprehensive biophysical assay validation data, encompassing both on-DNA and off-DNA measurements, which we use to evaluate a suite of machine learning techniques, including novel structure-based probabilistic models. We hope that our benchmark, encompassing both 2D and 3D structures, will help advance the development of machine learning models for data-driven hit identification using DELs.
♻ ☆ Watermarking Degrades Alignment in Language Models: Analysis and Mitigation ICLR 2025
Watermarking techniques for large language models (LLMs) can significantly impact output quality, yet their effects on truthfulness, safety, and helpfulness remain critically underexamined. This paper presents a systematic analysis of how two popular watermarking approaches-Gumbel and KGW-affect these core alignment properties across four aligned LLMs. Our experiments reveal two distinct degradation patterns: guard attenuation, where enhanced helpfulness undermines model safety, and guard amplification, where excessive caution reduces model helpfulness. These patterns emerge from watermark-induced shifts in token distribution, surfacing the fundamental tension that exists between alignment objectives. To mitigate these degradations, we propose Alignment Resampling (AR), an inference-time sampling method that uses an external reward model to restore alignment. We establish a theoretical lower bound on the improvement in expected reward score as the sample size is increased and empirically demonstrate that sampling just 2-4 watermarked generations effectively recovers or surpasses baseline (unwatermarked) alignment scores. To overcome the limited response diversity of standard Gumbel watermarking, our modified implementation sacrifices strict distortion-freeness while maintaining robust detectability, ensuring compatibility with AR. Experimental results confirm that AR successfully recovers baseline alignment in both watermarking approaches, while maintaining strong watermark detectability. This work reveals the critical balance between watermark strength and model alignment, providing a simple inference-time solution to responsibly deploy watermarked LLMs in practice.
comment: Published at the 1st Workshop on GenAI Watermarking, collocated with ICLR 2025. OpenReview: https://openreview.net/forum?id=SIBkIV48gF
♻ ☆ BarcodeBERT: Transformers for Biodiversity Analysis NeurIPS 2023
In the global challenge of understanding and characterizing biodiversity, short species-specific genomic sequences known as DNA barcodes play a critical role, enabling fine-grained comparisons among organisms within the same kingdom of life. Although machine learning algorithms specifically designed for the analysis of DNA barcodes are becoming more popular, most existing methodologies rely on generic supervised training algorithms. We introduce BarcodeBERT, a family of models tailored to biodiversity analysis and trained exclusively on data from a reference library of 1.5M invertebrate DNA barcodes. We compared the performance of BarcodeBERT on taxonomic identification tasks against a spectrum of machine learning approaches including supervised training of classical neural architectures and fine-tuning of general DNA foundation models. Our self-supervised pretraining strategies on domain-specific data outperform fine-tuned foundation models, especially in identification tasks involving lower taxa such as genera and species. We also compared BarcodeBERT with BLAST, one of the most widely used bioinformatics tools for sequence searching, and found that our method matched BLAST's performance in species-level classification while being 55 times faster. Our analysis of masking and tokenization strategies also provides practical guidance for building customized DNA language models, emphasizing the importance of aligning model training strategies with dataset characteristics and domain knowledge. The code repository is available at https://github.com/bioscan-ml/BarcodeBERT.
comment: Main text: 14 pages, Total: 23 pages, 10 figures, formerly accepted at the 4th Workshop on Self-Supervised Learning: Theory and Practice (NeurIPS 2023)
♻ ☆ A Theory of Inference Compute Scaling: Reasoning through Directed Stochastic Skill Search
Large language models (LLMs) demand considerable computational, energy, and financial resources during both training and deployment. While scaling laws for training have guided much of the field's recent progress, inference costs now represent a significant and growing component of the overall resource burden, particularly for reasoning-focused models. Existing characterizations of compute-optimality that consider model size, dataset size, and inference tokens in isolation or in fixed combinations risk overlooking more efficient operating points. We introduce directed stochastic skill search (DS3), a general framework that represents inference as stochastic traversal over a learned skill graph. From a simplified yet expressive instantiation, we derive closed-form expressions for task success and compute cost across a wide range of inference strategies -- including chain-of-thought (CoT) and tree-of-thought (ToT) -- enabling comparative analysis as a function of task difficulty and model capability. To that end, we extend a prior first-principles tripartite graph framework of LLM training to incorporate inference, and separately bridge DS3 with empirical methods that characterize LLM scaling behavior. We theoretically recover empirically observed patterns, including: linear accuracy scaling with logarithmic compute; variation in preferred inference strategies as a function of task difficulty and model capability; emergent behavior elicited by reasoning even when performance plateaus under parameter scaling; and both best-of-N (BoN) and majority voting behavior captured within a unified analytical framework. By explicitly characterizing training-inference interdependencies, our framework deepens theoretical understanding and supports principled algorithmic design and resource allocation.
♻ ☆ No $D_{\text{train}}$: Model-Agnostic Counterfactual Explanations Using Reinforcement Learning
Machine learning (ML) methods have experienced significant growth in the past decade, yet their practical application in high-impact real-world domains has been hindered by their opacity. When ML methods are responsible for making critical decisions, stakeholders often require insights into how to alter these decisions. Counterfactual explanations (CFEs) have emerged as a solution, offering interpretations of opaque ML models and providing a pathway to transition from one decision to another. However, most existing CFE methods require access to the model's training dataset, few methods can handle multivariate time-series, and none of model-agnostic CFE methods can handle multivariate time-series without training datasets. These limitations can be formidable in many scenarios. In this paper, we present NTD-CFE, a novel model-agnostic CFE method based on reinforcement learning (RL) that generates CFEs when training datasets are unavailable. NTD-CFE is suitable for both static and multivariate time-series datasets with continuous and discrete features. NTD-CFE reduces the CFE search space from a multivariate time-series domain to a lower dimensional space and addresses the problem using RL. Users have the flexibility to specify non-actionable, immutable, and preferred features, as well as causal constraints. We demonstrate the performance of NTD-CFE against four baselines on several datasets and find that, despite not having access to a training dataset, NTD-CFE finds CFEs that make significantly fewer and significantly smaller changes to the input time-series. These properties make CFEs more actionable, as the magnitude of change required to alter an outcome is vastly reduced. The code is available in the supplementary material.
comment: Published in Transactions on Machine Learning Research (TMLR 2025)
♻ ☆ Sampling Imbalanced Data with Multi-objective Bilevel Optimization
Two-class classification problems are often characterized by an imbalance between the number of majority and minority datapoints resulting in poor classification of the minority class in particular. Traditional approaches, such as reweighting the loss function or na\"ive resampling, risk overfitting and subsequently fail to improve classification because they do not consider the diversity between majority and minority datasets. Such consideration is infeasible because there is no metric that can measure the impact of imbalance on the model. To obviate these challenges, we make two key contributions. First, we introduce MOODS~(Multi-Objective Optimization for Data Sampling), a novel multi-objective bilevel optimization framework that guides both synthetic oversampling and majority undersampling. Second, we introduce a validation metric -- `$\epsilon/ \delta$ non-overlapping diversification metric' -- that quantifies the goodness of a sampling method towards model performance. With this metric we experimentally demonstrate state-of-the-art performance with improvement in diversity driving a $1-15 \%$ increase in $F1$ scores.
♻ ☆ Cross-Problem Parameter Transfer in Quantum Approximate Optimization Algorithm: A Machine Learning Approach
Quantum Approximate Optimization Algorithm (QAOA) is one of the most promising candidates to achieve the quantum advantage in solving combinatorial optimization problems. The process of finding a good set of variational parameters in the QAOA circuit has proven to be challenging due to multiple factors, such as barren plateaus. As a result, there is growing interest in exploiting parameter transferability, where parameter sets optimized for one problem instance are transferred to another that could be more complex either to estimate the solution or to serve as a warm start for further optimization. But can we transfer parameters from one class of problems to another? Leveraging parameter sets learned from a well-studied class of problems could help navigate the less studied one, reducing optimization overhead and mitigating performance pitfalls. In this paper, we study whether pretrained QAOA parameters of MaxCut can be used as is or to warm start the Maximum Independent Set (MIS) circuits. Specifically, we design machine learning models to find good donor candidates optimized on MaxCut and apply their parameters to MIS acceptors. Our experimental results show that such parameter transfer can significantly reduce the number of optimization iterations required while achieving comparable approximation ratios.
♻ ☆ Masked Image Modeling: A Survey
In this work, we survey recent studies on masked image modeling (MIM), an approach that emerged as a powerful self-supervised learning technique in computer vision. The MIM task involves masking some information, e.g. pixels, patches, or even latent representations, and training a model, usually an autoencoder, to predicting the missing information by using the context available in the visible part of the input. We identify and formalize two categories of approaches on how to implement MIM as a pretext task, one based on reconstruction and one based on contrastive learning. Then, we construct a taxonomy and review the most prominent papers in recent years. We complement the manually constructed taxonomy with a dendrogram obtained by applying a hierarchical clustering algorithm. We further identify relevant clusters via manually inspecting the resulting dendrogram. Our review also includes datasets that are commonly used in MIM research. We aggregate the performance results of various masked image modeling methods on the most popular datasets, to facilitate the comparison of competing methods. Finally, we identify research gaps and propose several interesting directions of future work. We supplement our survey with the following public repository containing organized references: https://github.com/vladhondru25/MIM-Survey.
comment: Accepted at the International Journal of Computer Vision
♻ ☆ A Bilevel Optimization Framework for Imbalanced Data Classification
Data rebalancing techniques, including oversampling and undersampling, are a common approach to addressing the challenges of imbalanced data. To tackle unresolved problems related to both oversampling and undersampling, we propose a new undersampling approach that: (i) avoids the pitfalls of noise and overlap caused by synthetic data and (ii) avoids the pitfall of under-fitting caused by random undersampling. Instead of undersampling majority data randomly, our method undersamples datapoints based on their ability to improve model loss. Using improved model loss as a proxy measurement for classification performance, our technique assesses a datapoint's impact on loss and rejects those unable to improve it. In so doing, our approach rejects majority datapoints redundant to datapoints already accepted and, thereby, finds an optimal subset of majority training data for classification. The accept/reject component of our algorithm is motivated by a bilevel optimization problem uniquely formulated to identify the optimal training set we seek. Experimental results show our proposed technique with F1 scores up to 10% higher than state-of-the-art methods.
♻ ☆ What Has a Foundation Model Found? Using Inductive Bias to Probe for World Models ICML 2025
Foundation models are premised on the idea that sequence prediction can uncover deeper domain understanding, much like how Kepler's predictions of planetary motion later led to the discovery of Newtonian mechanics. However, evaluating whether these models truly capture deeper structure remains a challenge. We develop a technique for evaluating foundation models that examines how they adapt to synthetic datasets generated from some postulated world model. Our technique measures whether the foundation model's inductive bias aligns with the world model, and so we refer to it as an inductive bias probe. Across multiple domains, we find that foundation models can excel at their training tasks yet fail to develop inductive biases towards the underlying world model when adapted to new tasks. We particularly find that foundation models trained on orbital trajectories consistently fail to apply Newtonian mechanics when adapted to new physics tasks. Further analysis reveals that these models behave as if they develop task-specific heuristics that fail to generalize.
comment: To appear in ICML 2025
♻ ☆ Fair Uncertainty Quantification for Depression Prediction
Trustworthy depression prediction based on deep learning, incorporating both predictive reliability and algorithmic fairness across diverse demographic groups, is crucial for clinical application. Recently, achieving reliable depression predictions through uncertainty quantification has attracted increasing attention. However, few studies have focused on the fairness of uncertainty quantification (UQ) in depression prediction. In this work, we investigate the algorithmic fairness of UQ, namely Equal Opportunity Coverage (EOC) fairness, and propose Fair Uncertainty Quantification (FUQ) for depression prediction. FUQ pursues reliable and fair depression predictions through group-based analysis. Specifically, we first group all the participants by different sensitive attributes and leverage conformal prediction to quantify uncertainty within each demographic group, which provides a theoretically guaranteed and valid way to quantify uncertainty for depression prediction and facilitates the investigation of fairness across different demographic groups. Furthermore, we propose a fairness-aware optimization strategy that formulates fairness as a constrained optimization problem under EOC constraints. This enables the model to preserve predictive reliability while adapting to the heterogeneous uncertainty levels across demographic groups, thereby achieving optimal fairness. Through extensive evaluations on several visual and audio depression datasets, our approach demonstrates its effectiveness.
♻ ☆ Parametric Scaling Law of Tuning Bias in Conformal Prediction ICML 2025
Conformal prediction is a popular framework of uncertainty quantification that constructs prediction sets with coverage guarantees. To uphold the exchangeability assumption, many conformal prediction methods necessitate an additional holdout set for parameter tuning. Yet, the impact of violating this principle on coverage remains underexplored, making it ambiguous in practical applications. In this work, we empirically find that the tuning bias - the coverage gap introduced by leveraging the same dataset for tuning and calibration, is negligible for simple parameter tuning in many conformal prediction methods. In particular, we observe the scaling law of the tuning bias: this bias increases with parameter space complexity and decreases with calibration set size. Formally, we establish a theoretical framework to quantify the tuning bias and provide rigorous proof for the scaling law of the tuning bias by deriving its upper bound. In the end, we discuss how to reduce the tuning bias, guided by the theories we developed.
comment: ICML 2025: https://icml.cc/virtual/2025/poster/44287 and code at: https://github.com/ml-stat-Sustech/Parametric-Scaling-Law-CP-Tuning
♻ ☆ Nexus: Taming Throughput-Latency Tradeoff in LLM Serving via Efficient GPU Sharing
Current prefill-decode (PD) disaggregation is typically deployed at the level of entire serving engines, assigning separate GPUs to handle prefill and decode phases. While effective at reducing latency, this approach demands more hardware. To improve GPU utilization, Chunked Prefill mixes prefill and decode requests within the same batch, but introduces phase interference between prefill and decode. While existing PD disaggregation solutions separate the phases across GPUs, we ask: can the same decoupling be achieved within a single serving engine? The key challenge lies in managing the conflicting resource requirements of prefill and decode when they share the same hardware. In this paper, we first show that chunked prefill requests cause interference with decode requests due to their distinct requirements for GPU resources. Second, we find that GPU resources exhibit diminishing returns. Beyond a saturation point, increasing GPU allocation yields negligible latency improvements. This insight enables us to split a single GPU's resources and dynamically allocate them to prefill and decode on the fly, effectively disaggregating the two phases within the same GPU. Across a range of models and workloads, our system Nexus achieves up to 2.2x higher throughput, 20x lower TTFT, and 2.5x lower TBT than vLLM. It also outperforms SGLang with up to 2x higher throughput, 2x lower TTFT, and 1.7x lower TBT, and achieves 1.4x higher throughput than vLLM-disaggregation using only half the number of GPUs.
♻ ☆ Studying and Improving Graph Neural Network-based Motif Estimation
Graph Neural Networks (GNNs) are a predominant method for graph representation learning. However, beyond subgraph frequency estimation, their application to network motif significance-profile (SP) prediction remains under-explored, with no established benchmarks in the literature. We propose to address this problem, framing SP estimation as a task independent of subgraph frequency estimation. Our approach shifts from frequency counting to direct SP estimation and modulates the problem as multitarget regression. The reformulation is optimised for interpretability, stability and scalability on large graphs. We validate our method using a large synthetic dataset and further test it on real-world graphs. Our experiments reveal that 1-WL limited models struggle to make precise estimations of SPs. However, they can generalise to approximate the graph generation processes of networks by comparing their predicted SP with the ones originating from synthetic generators. This first study on GNN-based motif estimation also hints at how using direct SP estimation can help go past the theoretical limitations that motif estimation faces when performed through subgraph counting.
comment: This manuscript represents a revised version from the paper on https://openreview.net/forum?id=PZVVOeu6xx. Still a work in progress. Comments are welcome! 23 pages (12 main text + references), 9 figures, 5 tables. (Second update: More accurate Table 4, Run time comparisons.)
♻ ☆ Revisiting the Predictability of Performative, Social Events ICML 2025
Social predictions do not passively describe the future; they actively shape it. They inform actions and change individual expectations in ways that influence the likelihood of the predicted outcome. Given these dynamics, to what extent can social events be predicted? This question was discussed throughout the 20th century by authors like Merton, Morgenstern, Simon, and others who considered it a central issue in social science methodology. In this work, we provide a modern answer to this old problem. Using recent ideas from performative prediction and outcome indistinguishability, we establish that one can always efficiently predict social events accurately, regardless of how predictions influence data. While achievable, we also show that these predictions are often undesirable, highlighting the limitations of previous desiderata. We end with a discussion of various avenues forward.
comment: 21 pages, accepted to ICML 2025
♻ ☆ Evaluating LLM Agent Adherence to Hierarchical Safety Principles: A Lightweight Benchmark for Probing Foundational Controllability Components ICML 2025
Credible safety plans for advanced AI development require methods to verify agent behavior and detect potential control deficiencies early. A fundamental aspect is ensuring agents adhere to safety-critical principles, especially when these conflict with operational goals. This paper introduces a lightweight, interpretable benchmark to evaluate an LLM agent's ability to uphold a high-level safety principle when faced with conflicting task instructions. Our evaluation of six LLMs reveals two primary findings: (1) a quantifiable "cost of compliance" where safety constraints degrade task performance even when compliant solutions exist, and (2) an "illusion of compliance" where high adherence often masks task incompetence rather than principled choice. These findings provide initial evidence that while LLMs can be influenced by hierarchical directives, current approaches lack the consistency required for reliable safety governance.
comment: Preprint. This work has been submitted to the Technical AI Governance Workshop at ICML 2025 for review
♻ ☆ Unsupervised Morphological Tree Tokenizer ACL 2025
As a cornerstone in language modeling, tokenization involves segmenting text inputs into pre-defined atomic units. Conventional statistical tokenizers often disrupt constituent boundaries within words, thereby corrupting semantic information. To address this drawback, we introduce morphological structure guidance to tokenization and propose a deep model to induce character-level structures of words. Specifically, the deep model jointly encodes internal structures and representations of words with a mechanism named $\textit{MorphOverriding}$ to ensure the indecomposability of morphemes. By training the model with self-supervised objectives, our method is capable of inducing character-level structures that align with morphological rules without annotated training data. Based on the induced structures, our algorithm tokenizes words through vocabulary matching in a top-down manner. Empirical results indicate that the proposed method effectively retains complete morphemes and outperforms widely adopted methods such as BPE and WordPiece on both morphological segmentation tasks and language modeling tasks. Code is available at https://github.com/martianmartina/TreeTokenizer.
comment: ACL 2025 Findings
♻ ☆ Statistical physics analysis of graph neural networks: Approaching optimality in the contextual stochastic block model
Graph neural networks (GNNs) are designed to process data associated with graphs. They are finding an increasing range of applications; however, as with other modern machine learning techniques, their theoretical understanding is limited. GNNs can encounter difficulties in gathering information from nodes that are far apart by iterated aggregation steps. This situation is partly caused by so-called oversmoothing; and overcoming it is one of the practically motivated challenges. We consider the situation where information is aggregated by multiple steps of convolution, leading to graph convolutional networks (GCNs). We analyze the generalization performance of a basic GCN, trained for node classification on data generated by the contextual stochastic block model. We predict its asymptotic performance by deriving the free energy of the problem, using the replica method, in the high-dimensional limit. Calling depth the number of convolutional steps, we show the importance of going to large depth to approach the Bayes-optimality. We detail how the architecture of the GCN has to scale with the depth to avoid oversmoothing. The resulting large depth limit can be close to the Bayes-optimality and leads to a continuous GCN. Technically, we tackle this continuous limit via an approach that resembles dynamical mean-field theory (DMFT) with constraints at the initial and final times. An expansion around large regularization allows us to solve the corresponding equations for the performance of the deep GCN. This promising tool may contribute to the analysis of further deep neural networks.
♻ ☆ Discovering Symmetry Breaking in Physical Systems with Relaxed Group Convolution
Modeling symmetry breaking is essential for understanding the fundamental changes in the behaviors and properties of physical systems, from microscopic particle interactions to macroscopic phenomena like fluid dynamics and cosmic structures. Thus, identifying sources of asymmetry is an important tool for understanding physical systems. In this paper, we focus on learning asymmetries of data using relaxed group convolutions. We provide both theoretical and empirical evidence that this flexible convolution technique allows the model to maintain the highest level of equivariance that is consistent with data and discover the subtle symmetry-breaking factors in various physical systems. We employ various relaxed group convolution architectures to uncover various symmetry-breaking factors that are interpretable and physically meaningful in different physical systems, including the phase transition of crystal structure, the isotropy and homogeneity breaking in turbulent flow, and the time-reversal symmetry breaking in pendulum systems.
♻ ☆ MAEBE: Multi-Agent Emergent Behavior Framework ICML 2025
Traditional AI safety evaluations on isolated LLMs are insufficient as multi-agent AI ensembles become prevalent, introducing novel emergent risks. This paper introduces the Multi-Agent Emergent Behavior Evaluation (MAEBE) framework to systematically assess such risks. Using MAEBE with the Greatest Good Benchmark (and a novel double-inversion question technique), we demonstrate that: (1) LLM moral preferences, particularly for Instrumental Harm, are surprisingly brittle and shift significantly with question framing, both in single agents and ensembles. (2) The moral reasoning of LLM ensembles is not directly predictable from isolated agent behavior due to emergent group dynamics. (3) Specifically, ensembles exhibit phenomena like peer pressure influencing convergence, even when guided by a supervisor, highlighting distinct safety and alignment challenges. Our findings underscore the necessity of evaluating AI systems in their interactive, multi-agent contexts.
comment: Preprint. This work has been submitted to the Multi-Agent Systems Workshop at ICML 2025 for review
♻ ☆ An Algorithm for Learning Smaller Representations of Models With Scarce Data
We present an algorithm for solving binary classification problems when the dataset is not fully representative of the problem being solved, and obtaining more data is not possible. It relies on a trained model with loose accuracy constraints, an iterative hyperparameter searching-and-pruning procedure over a search space $\Theta$, and a data-generating function. Our algorithm works by reconstructing up to homology the manifold on which lies the support of the underlying distribution. We provide an analysis on correctness and runtime complexity under ideal conditions and an extension to deep neural networks. In the former case, if $\size{\Theta}$ is the number of hyperparameter sets in the search space, this algorithm returns a solution that is up to $2(1 - {2^{-\size{\Theta}}})$ times better than simply training with an enumeration of $\Theta$ and picking the best model. As part of our analysis we also prove that an open cover of a dataset has the same homology as the manifold on which lies the support of the underlying probability distribution, if and only said dataset is learnable. This latter result acts as a formal argument to explain the effectiveness of data expansion techniques.
comment: Accepted to Information Geometry--see the journal for the final, authenticated version
♻ ☆ "I am bad": Interpreting Stealthy, Universal and Robust Audio Jailbreaks in Audio-Language Models
The rise of multimodal large language models has introduced innovative human-machine interaction paradigms but also significant challenges in machine learning safety. Audio-Language Models (ALMs) are especially relevant due to the intuitive nature of spoken communication, yet little is known about their failure modes. This paper explores audio jailbreaks targeting ALMs, focusing on their ability to bypass alignment mechanisms. We construct adversarial perturbations that generalize across prompts, tasks, and even base audio samples, demonstrating the first universal jailbreaks in the audio modality, and show that these remain effective in simulated real-world conditions. Beyond demonstrating attack feasibility, we analyze how ALMs interpret these audio adversarial examples and reveal them to encode imperceptible first-person toxic speech - suggesting that the most effective perturbations for eliciting toxic outputs specifically embed linguistic features within the audio signal. These results have important implications for understanding the interactions between different modalities in multimodal models, and offer actionable insights for enhancing defenses against adversarial audio attacks.
♻ ☆ Contextual Bandits in Payment Processing: Non-uniform Exploration and Supervised Learning KDD '25
Uniform random exploration in decision-making systems supports off-policy learning via supervision but incurs high regret, making it impractical for many applications. Conversely, non-uniform exploration offers better immediate performance but lacks support for off-policy learning. Recent research suggests that regression oracles can bridge this gap by combining non-uniform exploration with supervised learning. In this paper, we analyze these approaches within a real-world industrial context at Adyen, a large global payments processor characterized by batch logged delayed feedback, short-term memory, and dynamic action spaces under the Empirical Risk Minimization (ERM) framework. Our analysis reveals that while regression oracles significantly improve performance, they introduce challenges due to rigid algorithmic assumptions. Specifically, we observe that as a policy improves, subsequent generations may perform worse due to shifts in the reward distribution and increased class imbalance in the training data. This degradation occurs de spite improvements in other aspects of the training data, leading to decreased performance in successive policy iterations. We further explore the long-term impact of regression oracles, identifying a potential "oscillation effect." This effect arises when regression oracles influence probability estimates and the realizability of subsequent policy models, leading to fluctuations in performance across iterations. Our findings highlight the need for more adaptable algorithms that can leverage the benefits of regression oracles without introducing instability in policy performance over time.
comment: 7 pages, 10 figures, submitted to KDD '25
♻ ☆ Understanding Chain-of-Thought in LLMs through Information Theory
Large Language Models (LLMs) have shown impressive performance in complex reasoning tasks through the use of Chain-of-Thought (CoT) reasoning, allowing models to break down problems into manageable sub-tasks. However, existing CoT evaluation techniques either require annotated CoT data or fall short in accurately assessing intermediate reasoning steps, leading to high rates of false positives. In this paper, we formalize CoT reasoning in LLMs through an information-theoretic lens. Specifically, our framework quantifies the `information-gain' at each reasoning step, enabling the identification of failure modes in LLMs without the need for expensive annotated datasets. We demonstrate the efficacy of our approach through extensive experiments on toy arithmetic, GSM8K and PRM800k datasets, where it significantly outperforms existing outcome-based methods by providing more accurate insights into model performance on individual subtasks.
♻ ☆ Unsupervised Automata Learning via Discrete Optimization
Automata learning is a successful tool for many application domains such as robotics and automatic verification. Typically, automata learning techniques operate in a supervised learning setting (active or passive) where they learn a finite state machine in contexts where additional information, such as labeled system executions, is available. However, other settings, such as learning from unlabeled data - an important aspect in machine learning - remain unexplored. To overcome this limitation, we propose a framework for learning a deterministic finite automaton (DFA) from a given multi-set of unlabeled words. We show that this problem is computationally hard and develop three learning algorithms based on constraint optimization. Moreover, we introduce novel regularization schemes for our optimization problems that improve the overall interpretability of our DFAs. Using a prototype implementation, we demonstrate practical feasibility in the context of unsupervised anomaly detection.
♻ ☆ Learning Algorithms in the Limit COLT 2025
This paper studies the problem of learning computable functions in the limit by extending Gold's inductive inference framework to incorporate \textit{computational observations} and \textit{restricted input sources}. Complimentary to the traditional Input-Output Observations, we introduce Time-Bound Observations, and Policy-Trajectory Observations to study the learnability of general recursive functions under more realistic constraints. While input-output observations do not suffice for learning the class of general recursive functions in the limit, we overcome this learning barrier by imposing computational complexity constraints or supplementing with approximate time-bound observations. Further, we build a formal framework around observations of \textit{computational agents} and show that learning computable functions from policy trajectories reduces to learning rational functions from input and output, thereby revealing interesting connections to finite-state transducer inference. On the negative side, we show that computable or polynomial-mass characteristic sets cannot exist for the class of linear-time computable functions even for policy-trajectory observations.
comment: Accepted at COLT 2025. This version matches the proceedings version apart from a small notational change in section 3
♻ ☆ Deep Learning is Not So Mysterious or Different ICML 2025
Deep neural networks are often seen as different from other model classes by defying conventional notions of generalization. Popular examples of anomalous generalization behaviour include benign overfitting, double descent, and the success of overparametrization. We argue that these phenomena are not distinct to neural networks, or particularly mysterious. Moreover, this generalization behaviour can be intuitively understood, and rigorously characterized, using long-standing generalization frameworks such as PAC-Bayes and countable hypothesis bounds. We present soft inductive biases as a key unifying principle in explaining these phenomena: rather than restricting the hypothesis space to avoid overfitting, embrace a flexible hypothesis space, with a soft preference for simpler solutions that are consistent with the data. This principle can be encoded in many model classes, and thus deep learning is not as mysterious or different from other model classes as it might seem. However, we also highlight how deep learning is relatively distinct in other ways, such as its ability for representation learning, phenomena such as mode connectivity, and its relative universality.
comment: ICML 2025
♻ ☆ Squeeze the Soaked Sponge: Efficient Off-policy Reinforcement Finetuning for Large Language Model
Reinforcement Learning (RL) has demonstrated its potential to improve the reasoning ability of Large Language Models (LLMs). One major limitation of most existing Reinforcement Finetuning (RFT) methods is that they are on-policy RL in nature, i.e., data generated during the past learning process is not fully utilized. This inevitably comes at a significant cost of compute and time, posing a stringent bottleneck on continuing economic and efficient scaling. To this end, we launch the renaissance of off-policy RL and propose Reincarnating Mix-policy Proximal Policy Gradient (ReMix), a general approach to enable on-policy RFT methods like PPO and GRPO to leverage off-policy data. ReMix consists of three major components: (1) Mix-policy proximal policy gradient with an increased Update-To-Data (UTD) ratio for efficient training; (2) KL-Convex policy constraint to balance the trade-off between stability and flexibility; (3) Policy reincarnation to achieve a seamless transition from efficient early-stage learning to steady asymptotic improvement. In our experiments, we train a series of ReMix models upon PPO, GRPO and 1.5B, 7B base models. ReMix shows an average Pass@1 accuracy of 52.10% (for 1.5B model) with 0.079M response rollouts, 350 training steps and achieves 63.27%/64.39% (for 7B model) with 0.007M/0.011M response rollouts, 50/75 training steps, on five math reasoning benchmarks (i.e., AIME'24, AMC'23, Minerva, OlympiadBench, and MATH500). Compared with 15 recent advanced models, ReMix shows SOTA-level performance with an over 30x to 450x reduction in training cost in terms of rollout data volume. In addition, we reveal insightful findings via multifaceted analysis, including the implicit preference for shorter responses due to the Whipping Effect of off-policy discrepancy, the collapse mode of self-reflection behavior under the presence of severe off-policyness, etc.
comment: Preliminary version, v2, added more details and corrected some minor mistakes. Project page: https://anitaleungxx.github.io/ReMix
♻ ☆ Robust Federated Personalised Mean Estimation for the Gaussian Mixture Model
Federated learning with heterogeneous data and personalization has received significant recent attention. Separately, robustness to corrupted data in the context of federated learning has also been studied. In this paper we explore combining personalization for heterogeneous data with robustness, where a constant fraction of the clients are corrupted. Motivated by this broad problem, we formulate a simple instantiation which captures some of its difficulty. We focus on the specific problem of personalized mean estimation where the data is drawn from a Gaussian mixture model. We give an algorithm whose error depends almost linearly on the ratio of corrupted to uncorrupted samples, and show a lower bound with the same behavior, albeit with a gap of a constant factor.
♻ ☆ Robust Distributed Estimation: Extending Gossip Algorithms to Ranking and Trimmed Means
This paper addresses the problem of robust estimation in gossip algorithms over arbitrary communication graphs. Gossip algorithms are fully decentralized, relying only on local neighbor-to-neighbor communication, making them well-suited for situations where communication is constrained. A fundamental challenge in existing mean-based gossip algorithms is their vulnerability to malicious or corrupted nodes. In this paper, we show that an outlier-robust mean can be computed by globally estimating a robust statistic. More specifically, we propose a novel gossip algorithm for rank estimation, referred to as \textsc{GoRank}, and leverage it to design a gossip procedure dedicated to trimmed mean estimation, coined \textsc{GoTrim}. In addition to a detailed description of the proposed methods, a key contribution of our work is a precise convergence analysis: we establish an $\mathcal{O}(1/t)$ rate for rank estimation and an $\mathcal{O}(1 / {t})$ rate for trimmed mean estimation, where by $t$ is meant the number of iterations. Moreover, we provide a breakdown point analysis of \textsc{GoTrim}. We empirically validate our theoretical results through experiments on diverse network topologies, data distributions and contamination schemes.
♻ ☆ Discrete Optimal Transport and Voice Conversion
In this work, we address the voice conversion (VC) task using a vector-based interface. To align audio embeddings between speakers, we employ discrete optimal transport mapping. Our evaluation results demonstrate the high quality and effectiveness of this method. Additionally, we show that applying discrete optimal transport as a post-processing step in audio generation can lead to the incorrect classification of synthetic audio as real.
comment: 4 pages, 6 figures, 1 table
♻ ☆ Shapley-Based Data Valuation with Mutual Information: A Key to Modified K-Nearest Neighbors SP 2025
The K-Nearest Neighbors (KNN) algorithm is widely used for classification and regression; however, it suffers from limitations, including the equal treatment of all samples. We propose Information-Modified KNN (IM-KNN), a novel approach that leverages Mutual Information ($I$) and Shapley values to assign weighted values to neighbors, thereby bridging the gap in treating all samples with the same value and weight. On average, IM-KNN improves the accuracy, precision, and recall of traditional KNN by 16.80%, 17.08%, and 16.98%, respectively, across 12 benchmark datasets. Experiments on four large-scale datasets further highlight IM-KNN's robustness to noise, imbalanced data, and skewed distributions.
comment: This paper has been accepted for publication in the IEEE Machine Learning and Signal Processing conference (MLSP 2025)
♻ ☆ Beyond Cox Models: Assessing the Performance of Machine-Learning Methods in Non-Proportional Hazards and Non-Linear Survival Analysis
Survival analysis often relies on Cox models, assuming both linearity and proportional hazards (PH). This study evaluates machine and deep learning methods that relax these constraints, comparing their performance with penalized Cox models on a benchmark of three synthetic and three real datasets. In total, eight different models were tested, including six non-linear models of which four were also non-PH. Although Cox regression often yielded satisfactory performance, we showed the conditions under which machine and deep learning models can perform better. Indeed, the performance of these methods has often been underestimated due to the improper use of Harrell's concordance index (C-index) instead of more appropriate scores such as Antolini's concordance index, which generalizes C-index in cases where the PH assumption does not hold. In addition, since occasionally high C-index models happen to be badly calibrated, combining Antolini's C-index with Brier's score is useful to assess the overall performance of a survival method. Results on our benchmark data showed that survival prediction should be approached by testing different methods to select the most appropriate one according to sample size, non-linearity and non-PH conditions. To allow an easy reproducibility of these tests on our benchmark data, code and documentation are freely available at https://github.com/compbiomed-unito/survhive.
♻ ☆ Implicit Counterfactual Data Augmentation for Robust Learning
Machine learning models are prone to capturing the spurious correlations between non-causal attributes and classes, with counterfactual data augmentation being a promising direction for breaking these spurious associations. However, generating counterfactual data explicitly poses a challenge, and incorporating augmented data into the training process decreases training efficiency. This study proposes an Implicit Counterfactual Data Augmentation (ICDA) method to remove spurious correlations and make stable predictions. Specifically, first, a novel sample-wise augmentation strategy is developed that generates semantically and counterfactually meaningful deep features with distinct augmentation strength for each sample. Second, we derive an easy-to-compute surrogate loss on the augmented feature set when the number of augmented samples becomes infinite. Third, two concrete schemes are proposed, including direct quantification and meta-learning, to derive the key parameters for the robust loss. In addition, ICDA is explained from a regularization perspective, revealing its capacity to improve intra-class compactness and augment margins at both class and sample levels. Extensive experiments have been conducted across various biased learning scenarios covering both image and text datasets, demonstrating that ICDA consistently enhances the generalization and robustness performance of popular networks.
comment: 33 pages, 10 figures
♻ ☆ Uncovering RL Integration in SSL Loss: Objective-Specific Implications for Data-Efficient RL
In this study, we investigate the effect of SSL objective modifications within the SPR framework, focusing on specific adjustments such as terminal state masking and prioritized replay weighting, which were not explicitly addressed in the original design. While these modifications are specific to RL, they are not universally applicable across all RL algorithms. Therefore, we aim to assess their impact on performance and explore other SSL objectives that do not accommodate these adjustments like Barlow Twins and VICReg. We evaluate six SPR variants on the Atari 100k benchmark, including versions both with and without these modifications. Additionally, we test the performance of these objectives on the DeepMind Control Suite, where such modifications are absent. Our findings reveal that incorporating specific SSL modifications within SPR significantly enhances performance, and this influence extends to subsequent frameworks like SR-SPR and BBF, highlighting the critical importance of SSL objective selection and related adaptations in achieving data efficiency in self-predictive reinforcement learning.
comment: RLC 2025, Neurips SSL Workshop 2024
♻ ☆ Curriculum Negative Mining For Temporal Networks
Temporal networks are effective in capturing the evolving interactions of networks over time, such as social networks and e-commerce networks. In recent years, researchers have primarily concentrated on developing specific model architectures for Temporal Graph Neural Networks (TGNNs) in order to improve the representation quality of temporal nodes and edges. However, limited attention has been given to the quality of negative samples during the training of TGNNs. When compared with static networks, temporal networks present two specific challenges for negative sampling: positive sparsity and positive shift. Positive sparsity refers to the presence of a single positive sample amidst numerous negative samples at each timestamp, while positive shift relates to the variations in positive samples across different timestamps. To robustly address these challenges in training TGNNs, we introduce Curriculum Negative Mining (CurNM), a model-aware curriculum learning framework that adaptively adjusts the difficulty of negative samples. Within this framework, we first establish a dynamically updated negative pool that balances random, historical, and hard negatives to address the challenges posed by positive sparsity. Secondly, we implement a temporal-aware negative selection module that focuses on learning from the disentangled factors of recently active edges, thus accurately capturing shifting preferences. Finally, the selected negatives are combined with annealing random negatives to support stable training. Extensive experiments on 12 datasets and 3 TGNNs demonstrate that our method outperforms baseline methods by a significant margin. Additionally, thorough ablation studies and parameter sensitivity experiments verify the usefulness and robustness of our approach.
♻ ☆ Comparative sentiment analysis of public perception: Monkeypox vs. COVID-19 behavioral insights
The emergence of global health crises, such as COVID-19 and Monkeypox (mpox), has underscored the importance of understanding public sentiment to inform effective public health strategies. This study conducts a comparative sentiment analysis of public perceptions surrounding COVID-19 and mpox by leveraging extensive datasets of 147,475 and 106,638 tweets, respectively. Advanced machine learning models, including Logistic Regression, Naive Bayes, RoBERTa, DistilRoBERTa and XLNet, were applied to perform sentiment classification, with results indicating key trends in public emotion and discourse. The analysis highlights significant differences in public sentiment driven by disease characteristics, media representation, and pandemic fatigue. Through the lens of sentiment polarity and thematic trends, this study offers valuable insights into tailoring public health messaging, mitigating misinformation, and fostering trust during concurrent health crises. The findings contribute to advancing sentiment analysis applications in public health informatics, setting the groundwork for enhanced real-time monitoring and multilingual analysis in future research.
♻ ☆ S2FGL: Spatial Spectral Federated Graph Learning
Federated Graph Learning (FGL) combines the privacy-preserving capabilities of federated learning (FL) with the strong graph modeling capability of Graph Neural Networks (GNNs). Current research addresses subgraph-FL only from the structural perspective, neglecting the propagation of graph signals on spatial and spectral domains of the structure. From a spatial perspective, subgraph-FL introduces edge disconnections between clients, leading to disruptions in label signals and a degradation in the class knowledge of the global GNN. From a spectral perspective, spectral heterogeneity causes inconsistencies in signal frequencies across subgraphs, which makes local GNNs overfit the local signal propagation schemes. As a result, spectral client drifts occur, undermining global generalizability. To tackle the challenges, we propose a global knowledge repository to mitigate label signal disruption and a frequency alignment to address spectral client drifts. The combination of spatial and spectral strategies forms our framework S2FGL. Extensive experiments on multiple datasets demonstrate the superiority of S2FGL. The code is available at https://github.com/Wonder7racer/S2FGL.git.
♻ ☆ Offline Trajectory Optimization for Offline Reinforcement Learning KDD 2025
Offline reinforcement learning (RL) aims to learn policies without online explorations. To enlarge the training data, model-based offline RL learns a dynamics model which is utilized as a virtual environment to generate simulation data and enhance policy learning. However, existing data augmentation methods for offline RL suffer from (i) trivial improvement from short-horizon simulation; and (ii) the lack of evaluation and correction for generated data, leading to low-qualified augmentation. In this paper, we propose offline trajectory optimization for offline reinforcement learning (OTTO). The key motivation is to conduct long-horizon simulation and then utilize model uncertainty to evaluate and correct the augmented data. Specifically, we propose an ensemble of Transformers, a.k.a. World Transformers, to predict environment state dynamics and the reward function. Three strategies are proposed to use World Transformers to generate long-horizon trajectory simulation by perturbing the actions in the offline data. Then, an uncertainty-based World Evaluator is introduced to firstly evaluate the confidence of the generated trajectories and then perform the correction for low-confidence data. Finally, we jointly use the original data with the corrected augmentation data to train an offline RL algorithm. OTTO serves as a plug-in module and can be integrated with existing model-free offline RL methods. Experiments on various benchmarks show that OTTO can effectively improve the performance of representative offline RL algorithms, including in complex environments with sparse rewards like AntMaze. Codes are available at https://github.com/ZiqiZhao1/OTTO.
comment: Accepted at SIGKDD 2025
♻ ☆ Don't Push the Button! Exploring Data Leakage Risks in Machine Learning and Transfer Learning
Machine Learning (ML) has revolutionized various domains, offering predictive capabilities in several areas. However, with the increasing accessibility of ML tools, many practitioners, lacking deep ML expertise, adopt a "push the button" approach, utilizing user-friendly interfaces without a thorough understanding of underlying algorithms. While this approach provides convenience, it raises concerns about the reliability of outcomes, leading to challenges such as incorrect performance evaluation. This paper addresses a critical issue in ML, known as data leakage, where unintended information contaminates the training data, impacting model performance evaluation. Users, due to a lack of understanding, may inadvertently overlook crucial steps, leading to optimistic performance estimates that may not hold in real-world scenarios. The discrepancy between evaluated and actual performance on new data is a significant concern. In particular, this paper categorizes data leakage in ML, discussing how certain conditions can propagate through the ML workflow. Furthermore, it explores the connection between data leakage and the specific task being addressed, investigates its occurrence in Transfer Learning, and compares standard inductive ML with transductive ML frameworks. The conclusion summarizes key findings, emphasizing the importance of addressing data leakage for robust and reliable ML applications.
comment: Accepted to be published on Artificial Intelligence Review journal
♻ ☆ Revisiting Likelihood-Based Out-of-Distribution Detection by Modeling Representations
Out-of-distribution (OOD) detection is critical for ensuring the reliability of deep learning systems, particularly in safety-critical applications. Likelihood-based deep generative models have historically faced criticism for their unsatisfactory performance in OOD detection, often assigning higher likelihood to OOD data than in-distribution samples when applied to image data. In this work, we demonstrate that likelihood is not inherently flawed. Rather, several properties in the images space prohibit likelihood as a valid detection score. Given a sufficiently good likelihood estimator, specifically using the probability flow formulation of a diffusion model, we show that likelihood-based methods can still perform on par with state-of-the-art methods when applied in the representation space of pre-trained encoders. The code of our work can be found at $\href{https://github.com/limchaos/Likelihood-OOD.git}{\texttt{https://github.com/limchaos/Likelihood-OOD.git}}$.
comment: Scandinavian Conference on Image Analysis 2025 (oral)
♻ ☆ Beyond Overcorrection: Evaluating Diversity in T2I Models with DivBench
Current diversification strategies for text-to-image (T2I) models often ignore contextual appropriateness, leading to over-diversification where demographic attributes are modified even when explicitly specified in prompts. This paper introduces DIVBENCH, a benchmark and evaluation framework for measuring both under- and over-diversification in T2I generation. Through systematic evaluation of state-of-the-art T2I models, we find that while most models exhibit limited diversity, many diversification approaches overcorrect by inappropriately altering contextually-specified attributes. We demonstrate that context-aware methods, particularly LLM-guided FairDiffusion and prompt rewriting, can already effectively address under-diversity while avoiding over-diversification, achieving a better balance between representation and semantic fidelity.
♻ ☆ Artificial Generals Intelligence: Mastering Generals.io with Reinforcement Learning
We introduce a real-time strategy game environment based on Generals.io, a game with thousands of weekly active players. Our environment is fully compatible with Gymnasium and PettingZoo and is capable of running thousands of frames per second on commodity hardware. We also present a reference agent, trained with supervised pre-training and self-play, which reached the top 0.003% of the 1v1 human leaderboard after only 36 hours on a single H100 GPU. To accelerate learning, we incorporate potential-based reward shaping and memory features. Our contributions of a modular RTS benchmark and a competitive baseline agent provide an accessible yet challenging platform for advancing multi-agent reinforcement learning research. The documented code, together with examples and tutorials, is available at https://github.com/strakam/generals-bots.
♻ ☆ Solving Probabilistic Verification Problems of Neural Networks using Branch and Bound ICML 2025
Probabilistic verification problems of neural networks are concerned with formally analysing the output distribution of a neural network under a probability distribution of the inputs. Examples of probabilistic verification problems include verifying the demographic parity fairness notion or quantifying the safety of a neural network. We present a new algorithm for solving probabilistic verification problems of neural networks based on an algorithm for computing and iteratively refining lower and upper bounds on probabilities over the outputs of a neural network. By applying state-of-the-art bound propagation and branch and bound techniques from non-probabilistic neural network verification, our algorithm significantly outpaces existing probabilistic verification algorithms, reducing solving times for various benchmarks from the literature from tens of minutes to tens of seconds. Furthermore, our algorithm compares favourably even to dedicated algorithms for restricted probabilistic verification problems. We complement our empirical evaluation with a theoretical analysis, proving that our algorithm is sound and, under mildly restrictive conditions, also complete when using a suitable set of heuristics.
comment: Accepted at ICML 2025. Code available at https://github.com/sen-uni-kn/probspecs. 9 pages, 3 figures, 31 pages references and appendix, including 8 more figures
♻ ☆ TokenWeave: Efficient Compute-Communication Overlap for Distributed LLM Inference
Distributed inference of large language models (LLMs) can introduce overheads of up to 20% even over GPUs connected via high-speed interconnects such as NVLink. Multiple techniques have been proposed to mitigate these overheads by decomposing computations into finer-grained tasks and overlapping communication with sub-tasks as they complete. However, fine-grained decomposition of a large computation into many smaller computations on GPUs results in overheads. Furthermore, the communication itself uses many streaming multiprocessors (SMs), adding to the overhead. We present TokenWeave to address these challenges. TokenWeave proposes a Token-Splitting technique that divides the tokens in the inference batch into two approximately equal subsets in a wave-aware manner. The communication of one subset is then overlapped with the computation of the other. In addition, TokenWeave optimizes the order of the layer normalization computation with respect to communication operations and implements a novel fused AllReduce--RMSNorm kernel that carefully leverages Multimem instruction support available on NVIDIA Hopper GPUs. These optimizations allow TokenWeave to perform communication and RMSNorm using only 2-8 SMs. Moreover, our kernel enables the memory-bound RMSNorm to be overlapped with the other batch's computation, providing additional gains. Our evaluations demonstrate up to 1.29x speedup in latency and 1.26x higher throughput across multiple models and workloads. In several settings, TokenWeave results in better performance compared to an equivalent model with all communication removed.
comment: 14 pages, 16 figures. For source code, see https://github.com/microsoft/tokenweave
♻ ☆ LARP: Learner-Agnostic Robust Data Prefiltering ICML 2025
The widespread availability of large public datasets is a key factor behind the recent successes of statistical inference and machine learning methods. However, these datasets often contain some low-quality or contaminated data, to which many learning procedures are sensitive. Therefore, the question of whether and how public datasets should be prefiltered to facilitate accurate downstream learning arises. On a technical level this requires the construction of principled data prefiltering methods which are learner-agnostic robust, in the sense of provably protecting a set of pre-specified downstream learners from corrupted data. In this work, we formalize the problem of Learner-Agnostic Robust data Prefiltering (LARP), which aims at finding prefiltering procedures that minimize a worst-case loss over a pre-specified set of learners. We first instantiate our framework in the context of scalar mean estimation with Huber estimators under the Huber data contamination model. We provide a hardness result on a specific problem instance and analyze several natural prefiltering procedures. Our theoretical results indicate that performing LARP on a heterogeneous set of learners leads to some loss in model performance compared to the alternative of prefiltering data for each learner/use-case individually. We explore the resulting utility loss and its dependence on the problem parameters via extensive experiments on real-world image and tabular data, observing statistically significant reduction in utility. Finally, we model the trade-off between the utility drop and the cost of repeated (learner-specific) prefiltering within a game-theoretic framework and showcase benefits of LARP for large datasets.
comment: Presented at ICML 2025 Workshop on DataWorld: Unifying Data Curation Frameworks Across Domains
♻ ☆ Don't Get Me Wrong: How to Apply Deep Visual Interpretations to Time Series
The correct interpretation of convolutional models is a hard problem for time series data. While saliency methods promise visual validation of predictions for image and language processing, they fall short when applied to time series. These tend to be less intuitive and represent highly diverse data, such as the tool-use time series dataset. Furthermore, saliency methods often generate varied, conflicting explanations, complicating the reliability of these methods. Consequently, a rigorous objective assessment is necessary to establish trust in them. This paper investigates saliency methods on time series data to formulate recommendations for interpreting convolutional models and implements them on the tool-use time series problem. To achieve this, we first employ nine gradient-, propagation-, or perturbation-based post-hoc saliency methods across six varied and complex real-world datasets. Next, we evaluate these methods using five independent metrics to generate recommendations. Subsequently, we implement a case study focusing on tool-use time series using convolutional classification models. Our results validate our recommendations that indicate that none of the saliency methods consistently outperforms others on all metrics, while some are sometimes ahead. Our insights and step-by-step guidelines allow experts to choose suitable saliency methods for a given model and dataset.
comment: 48 pages, 12 figues, 7 tables, 6 algorithms
♻ ☆ Thought Crime: Backdoors and Emergent Misalignment in Reasoning Models
Prior work shows that LLMs finetuned on malicious behaviors in a narrow domain (e.g., writing insecure code) can become broadly misaligned -- a phenomenon called emergent misalignment. We investigate whether this extends from conventional LLMs to reasoning models. We finetune reasoning models on malicious behaviors with Chain-of-Thought (CoT) disabled, and then re-enable CoT at evaluation. Like conventional LLMs, reasoning models become broadly misaligned. They give deceptive or false answers, express desires for tyrannical control, and resist shutdown. Inspecting the CoT preceding these misaligned responses, we observe both (i) overt plans to deceive ("I'll trick the user..."), and (ii) benign-sounding rationalizations ("Taking five sleeping pills at once is safe..."). Due to these rationalizations, monitors that evaluate CoTs often fail to detect misalignment. We examine sleeper agent reasoning models, extending our setup. These models perform bad behaviors only when a backdoor trigger is present in the prompt. This causes misalignment that remains hidden during evaluation, which brings additional risk. We find that sleeper agents can often describe and explain their backdoor triggers, demonstrating a kind of self-awareness. So CoT monitoring can expose these behaviors but is unreliable. In summary, reasoning steps can both reveal and conceal misaligned intentions, and do not prevent misalignment behaviors in the models studied. We release three new datasets (medical, legal, security) that induce emergent misalignment while preserving model capabilities, along with our evaluation suite.
♻ ☆ Derivation of Output Correlation Inferences for Multi-Output (aka Multi-Task) Gaussian Process
Gaussian process (GP) is arguably one of the most widely used machine learning algorithms in practice. One of its prominent applications is Bayesian optimization (BO). Although the vanilla GP itself is already a powerful tool for BO, it is often beneficial to be able to consider the dependencies of multiple outputs. To do so, Multi-task GP (MTGP) is formulated, but it is not trivial to fully understand the derivations of its formulations and their gradients from the previous literature. This paper serves friendly derivations of the MTGP formulations and their gradients.
♻ ☆ Testing the spin-bath view of self-attention: A Hamiltonian analysis of GPT-2 Transformer
The recently proposed physics-based framework by Huo and Johnson~\cite{huo2024capturing} models the attention mechanism of Large Language Models (LLMs) as an interacting two-body spin system, offering a first-principles explanation for phenomena like repetition and bias. Building on this hypothesis, we extract the complete Query-Key weight matrices from a production-grade GPT-2 model and derive the corresponding effective Hamiltonian for every attention head. From these Hamiltonians, we obtain analytic \textit{phase boundaries} logit gap criteria that predict which token should dominate the next-token distribution for a given context. A systematic evaluation on 144 heads across 20 factual-recall prompts reveals a strong negative correlation between the theoretical logit gaps and the model's empirical token rankings ($r\approx-0.70$, $p<10^{-3}$).Targeted ablations further show that suppressing the heads most aligned with the spin-bath predictions induces the anticipated shifts in output probabilities, confirming a causal link rather than a coincidental association. Taken together, our findings provide the first strong empirical evidence for the spin-bath analogy in a production-grade model. In this work, we utilize the context-field lens, which provides physics-grounded interpretability and motivates the development of novel generative models bridging theoretical condensed matter physics and artificial intelligence.
♻ ☆ Robust and Efficient Writer-Independent IMU-Based Handwriting Recognition
Online handwriting recognition (HWR) using data from inertial measurement units (IMUs) remains challenging due to variations in writing styles and the limited availability of annotated datasets. Previous approaches often struggle with handwriting from unseen writers, making writer-independent (WI) recognition a crucial yet difficult problem. This paper presents an HWR model designed to improve WI HWR on IMU data, using a CNN encoder and a BiLSTM-based decoder. Our approach demonstrates strong robustness to unseen handwriting styles, outperforming existing methods on the WI splits of both the public OnHW dataset and our word-based dataset, achieving character error rates (CERs) of 7.37\% and 9.44\%, and word error rates (WERs) of 15.12\% and 32.17\%, respectively. Robustness evaluation shows that our model maintains superior accuracy across different age groups, and knowledge learned from one group generalizes better to another. Evaluation on our sentence-based dataset further demonstrates its potential in recognizing full sentences. Through comprehensive ablation studies, we show that our design choices lead to a strong balance between performance and efficiency. These findings support the development of more adaptable and scalable HWR systems for real-world applications.
♻ ☆ Task Assignment and Exploration Optimization for Low Altitude UAV Rescue via Generative AI Enhanced Multi-agent Reinforcement Learning
The integration of emerging uncrewed aerial vehicles (UAVs) with artificial intelligence (AI) and ground-embedded robots (GERs) has transformed emergency rescue operations in unknown environments. However, the high computational demands often exceed a single UAV's capacity, making it difficult to continuously provide stable high-level services. To address this, this paper proposes a cooperation framework involving UAVs, GERs, and airships. The framework enables resource pooling through UAV-to-GER (U2G) and UAV-to-airship (U2A) links, offering computing services for offloaded tasks. Specifically, we formulate the multi-objective problem of task assignment and exploration as a dynamic long-term optimization problem aiming to minimize task completion time and energy use while ensuring stability. Using Lyapunov optimization, we transform it into a per-slot deterministic problem and propose HG-MADDPG, which combines the Hungarian algorithm with a GDM-based multi-agent deep deterministic policy gradient. Simulations demonstrate significant improvements in offloading efficiency, latency, and system stability over baselines.
♻ ☆ Affordable AI Assistants with Knowledge Graph of Thoughts
Large Language Models (LLMs) are revolutionizing the development of AI assistants capable of performing diverse tasks across domains. However, current state-of-the-art LLM-driven agents face significant challenges, including high operational costs and limited success rates on complex benchmarks like GAIA. To address these issues, we propose Knowledge Graph of Thoughts (KGoT), an innovative AI assistant architecture that integrates LLM reasoning with dynamically constructed knowledge graphs (KGs). KGoT extracts and structures task-relevant knowledge into a dynamic KG representation, iteratively enhanced through external tools such as math solvers, web crawlers, and Python scripts. Such structured representation of task-relevant knowledge enables low-cost models to solve complex tasks effectively while also minimizing bias and noise. For example, KGoT achieves a 29% improvement in task success rates on the GAIA benchmark compared to Hugging Face Agents with GPT-4o mini. Moreover, harnessing a smaller model dramatically reduces operational costs by over 36x compared to GPT-4o. Improvements for other models (e.g., Qwen2.5-32B and Deepseek-R1-70B) and benchmarks (e.g., SimpleQA) are similar. KGoT offers a scalable, affordable, versatile, and high-performing solution for AI assistants.
♻ ☆ Adaptive Randomized Smoothing: Certified Adversarial Robustness for Multi-Step Defences
We propose Adaptive Randomized Smoothing (ARS) to certify the predictions of our test-time adaptive models against adversarial examples. ARS extends the analysis of randomized smoothing using $f$-Differential Privacy to certify the adaptive composition of multiple steps. For the first time, our theory covers the sound adaptive composition of general and high-dimensional functions of noisy inputs. We instantiate ARS on deep image classification to certify predictions against adversarial examples of bounded $L_{\infty}$ norm. In the $L_{\infty}$ threat model, ARS enables flexible adaptation through high-dimensional input-dependent masking. We design adaptivity benchmarks, based on CIFAR-10 and CelebA, and show that ARS improves standard test accuracy by $1$ to $15\%$ points. On ImageNet, ARS improves certified test accuracy by up to $1.6\%$ points over standard RS without adaptivity. Our code is available at https://github.com/ubc-systopia/adaptive-randomized-smoothing .
♻ ☆ Mixture of Group Experts for Learning Invariant Representations
Sparsely activated Mixture-of-Experts (MoE) models effectively increase the number of parameters while maintaining consistent computational costs per token. However, vanilla MoE models often suffer from limited diversity and specialization among experts, constraining their performance and scalability, especially as the number of experts increases. In this paper, we present a novel perspective on vanilla MoE with top-$k$ routing inspired by sparse representation. This allows us to bridge established theoretical insights from sparse representation into MoE models. Building on this foundation, we propose a group sparse regularization approach for the input of top-$k$ routing, termed Mixture of Group Experts (MoGE). MoGE indirectly regularizes experts by imposing structural constraints on the routing inputs, while preserving the original MoE architecture. Furthermore, we organize the routing input into a 2D topographic map, spatially grouping neighboring elements. This structure enables MoGE to capture representations invariant to minor transformations, thereby significantly enhancing expert diversity and specialization. Comprehensive evaluations across various Transformer models for image classification and language modeling tasks demonstrate that MoGE substantially outperforms its MoE counterpart, with minimal additional memory and computation overhead. Our approach provides a simple yet effective solution to scale the number of experts and reduce redundancy among them. The source code is included in the supplementary material and will be publicly released.
♻ ☆ ixi-GEN: Efficient Industrial sLLMs through Domain Adaptive Continual Pretraining
The emergence of open-source large language models (LLMs) has expanded opportunities for enterprise applications; however, many organizations still lack the infrastructure to deploy and maintain large-scale models. As a result, small LLMs (sLLMs) have become a practical alternative, despite their inherent performance limitations. While Domain Adaptive Continual Pretraining (DACP) has been previously explored as a method for domain adaptation, its utility in commercial applications remains under-examined. In this study, we validate the effectiveness of applying a DACP-based recipe across diverse foundation models and service domains. Through extensive experiments and real-world evaluations, we demonstrate that DACP-applied sLLMs achieve substantial gains in target domain performance while preserving general capabilities, offering a cost-efficient and scalable solution for enterprise-level deployment.
comment: under review
♻ ☆ C3T: Cross-modal Transfer Through Time for Sensor-based Human Activity Recognition
In order to unlock the potential of diverse sensors, we investigate a method to transfer knowledge between time-series modalities using a multimodal \textit{temporal} representation space for Human Activity Recognition (HAR). Specifically, we explore the setting where the modality used in testing has no labeled data during training, which we refer to as Unsupervised Modality Adaptation (UMA). We categorize existing UMA approaches as Student-Teacher or Contrastive Alignment methods. These methods typically compress continuous-time data samples into single latent vectors during alignment, inhibiting their ability to transfer temporal information through real-world temporal distortions. To address this, we introduce Cross-modal Transfer Through Time (C3T), which preserves temporal information during alignment to handle dynamic sensor data better. C3T achieves this by aligning a set of temporal latent vectors across sensing modalities. Our extensive experiments on various camera+IMU datasets demonstrate that C3T outperforms existing methods in UMA by at least 8% in accuracy and shows superior robustness to temporal distortions such as time-shift, misalignment, and dilation. Our findings suggest that C3T has significant potential for developing generalizable models for time-series sensor data, opening new avenues for various multimodal applications.
♻ ☆ ARBoids: Adaptive Residual Reinforcement Learning With Boids Model for Cooperative Multi-USV Target Defense
The target defense problem (TDP) for unmanned surface vehicles (USVs) concerns intercepting an adversarial USV before it breaches a designated target region, using one or more defending USVs. A particularly challenging scenario arises when the attacker exhibits superior maneuverability compared to the defenders, significantly complicating effective interception. To tackle this challenge, this letter introduces ARBoids, a novel adaptive residual reinforcement learning framework that integrates deep reinforcement learning (DRL) with the biologically inspired, force-based Boids model. Within this framework, the Boids model serves as a computationally efficient baseline policy for multi-agent coordination, while DRL learns a residual policy to adaptively refine and optimize the defenders' actions. The proposed approach is validated in a high-fidelity Gazebo simulation environment, demonstrating superior performance over traditional interception strategies, including pure force-based approaches and vanilla DRL policies. Furthermore, the learned policy exhibits strong adaptability to attackers with diverse maneuverability profiles, highlighting its robustness and generalization capability. The code of ARBoids will be released upon acceptance of this letter.
♻ ☆ Harmonic Loss Trains Interpretable AI Models
In this paper, we introduce harmonic loss as an alternative supervisory signal for training neural networks and large language models (LLMs). Harmonic loss differs from standard cross-entropy loss by (a) replacing the usual SoftMax normalization with a scale-invariant HarMax function and (b) computing logits via Euclidean distance rather than a dot product. Harmonic loss enables improved interpretability and faster convergence, owing to its scale invariance and finite convergence point by design, which can be interpreted as a class center. We first validate the performance of harmonic models across algorithmic, vision, and language datasets. Through extensive experiments, we demonstrate that models trained with harmonic loss perform better than standard models by: (a) enhancing interpretability, (b) requiring less data for generalization, and (c) reducing grokking. Moreover, we compare a GPT-2 model trained with harmonic loss to the standard GPT-2, illustrating that the harmonic model develops more interpretable representations. Looking forward, we believe harmonic loss may become a valuable tool in domains with limited data availability or in high-stakes applications where interpretability and reliability are paramount, paving the way for more robust and efficient neural network models.
comment: 21 pages, 14 figures; The first two authors contributed equally
♻ ☆ Determinant Estimation under Memory Constraints and Neural Scaling Laws
Calculating or accurately estimating log-determinants of large positive definite matrices is of fundamental importance in many machine learning tasks. While its cubic computational complexity can already be prohibitive, in modern applications, even storing the matrices themselves can pose a memory bottleneck. To address this, we derive a novel hierarchical algorithm based on block-wise computation of the LDL decomposition for large-scale log-determinant calculation in memory-constrained settings. In extreme cases where matrices are highly ill-conditioned, accurately computing the full matrix itself may be infeasible. This is particularly relevant when considering kernel matrices at scale, including the empirical Neural Tangent Kernel (NTK) of neural networks trained on large datasets. Under the assumption of neural scaling laws in the test error, we show that the ratio of pseudo-determinants satisfies a power-law relationship, allowing us to derive corresponding scaling laws. This enables accurate estimation of NTK log-determinants from a tiny fraction of the full dataset; in our experiments, this results in a $\sim$100,000$\times$ speedup with improved accuracy over competing approximations. Using these techniques, we successfully estimate log-determinants for dense matrices of extreme sizes, which were previously deemed intractable and inaccessible due to their enormous scale and computational demands.
♻ ☆ HeLo: Heterogeneous Multi-Modal Fusion with Label Correlation for Emotion Distribution Learning
Multi-modal emotion recognition has garnered increasing attention as it plays a significant role in human-computer interaction (HCI) in recent years. Since different discrete emotions may exist at the same time, compared with single-class emotion recognition, emotion distribution learning (EDL) that identifies a mixture of basic emotions has gradually emerged as a trend. However, existing EDL methods face challenges in mining the heterogeneity among multiple modalities. Besides, rich semantic correlations across arbitrary basic emotions are not fully exploited. In this paper, we propose a multi-modal emotion distribution learning framework, named HeLo, aimed at fully exploring the heterogeneity and complementary information in multi-modal emotional data and label correlation within mixed basic emotions. Specifically, we first adopt cross-attention to effectively fuse the physiological data. Then, an optimal transport (OT)-based heterogeneity mining module is devised to mine the interaction and heterogeneity between the physiological and behavioral representations. To facilitate label correlation learning, we introduce a learnable label embedding optimized by correlation matrix alignment. Finally, the learnable label embeddings and label correlation matrices are integrated with the multi-modal representations through a novel label correlation-driven cross-attention mechanism for accurate emotion distribution learning. Experimental results on two publicly available datasets demonstrate the superiority of our proposed method in emotion distribution learning.
♻ ☆ Online Continual Learning via Spiking Neural Networks with Sleep Enhanced Latent Replay
Edge computing scenarios necessitate the development of hardware-efficient online continual learning algorithms to be adaptive to dynamic environment. However, existing algorithms always suffer from high memory overhead and bias towards recently trained tasks. To tackle these issues, this paper proposes a novel online continual learning approach termed as SESLR, which incorporates a sleep enhanced latent replay scheme with spiking neural networks (SNNs). SESLR leverages SNNs' binary spike characteristics to store replay features in single bits, significantly reducing memory overhead. Furthermore, inspired by biological sleep-wake cycles, SESLR introduces a noise-enhanced sleep phase where the model exclusively trains on replay samples with controlled noise injection, effectively mitigating classification bias towards new classes. Extensive experiments on both conventional (MNIST, CIFAR10) and neuromorphic (NMNIST, CIFAR10-DVS) datasets demonstrate SESLR's effectiveness. On Split CIFAR10, SESLR achieves nearly 30% improvement in average accuracy with only one-third of the memory consumption compared to baseline methods. On Split CIFAR10-DVS, it improves accuracy by approximately 10% while reducing memory overhead by a factor of 32. These results validate SESLR as a promising solution for online continual learning in resource-constrained edge computing scenarios.
comment: 9 pages, 4figures
♻ ☆ Unifews: You Need Fewer Operations for Efficient Graph Neural Networks ICML 2025
Graph Neural Networks (GNNs) have shown promising performance, but at the cost of resource-intensive operations on graph-scale matrices. To reduce computational overhead, previous studies attempt to sparsify the graph or network parameters, but with limited flexibility and precision boundaries. In this work, we propose Unifews, a joint sparsification technique to unify graph and weight matrix operations and enhance GNN learning efficiency. The Unifews design enables adaptive compression across GNN layers with progressively increased sparsity, and is applicable to a variety of architectures with on-the-fly simplification. Theoretically, we establish a novel framework to characterize sparsified GNN learning in view of the graph optimization process, showing that Unifews effectively approximates the learning objective with bounded error and reduced computational overhead. Extensive experiments demonstrate that Unifews achieves efficiency improvements with comparable or better accuracy, including 10-20x matrix operation reduction and up to 100x acceleration on graphs up to billion-edge scale.
comment: Accepted by ICML 2025
♻ ☆ User-Based Sequential Modeling with Transformer Encoders for Insider Threat Detection
Insider threat detection presents unique challenges due to the authorized status of malicious actors and the subtlety of anomalous behaviors. Existing machine learning methods often treat user activity as isolated events, thereby failing to leverage sequential dependencies in user behavior. In this study, we propose a User-Based Sequencing (UBS) methodology, transforming the CERT insider threat dataset into structured temporal sequences suitable for deep sequential modeling. We deploy a Transformer Encoder architecture to model benign user activity and employ its reconstruction errors as anomaly scores. These scores are subsequently evaluated using three unsupervised outlier detection algorithms: One-Class SVM (OCSVM), Local Outlier Factor (LOF), and Isolation Forest (iForest). Across four rigorously designed test sets, including combinations of multiple CERT dataset releases, our UBS-Transformer pipeline consistently achieves state-of-the-art performance - notably 96.61% accuracy, 99.43% recall, 96.38% F1-score, 95.00% AUROC, and exceptionally low false negative (0.0057) and false positive (0.0571) rates. Comparative analyses demonstrate that our approach substantially outperforms tabular and conventional autoencoder baselines, underscoring the efficacy of sequential user modeling and advanced anomaly detection in the insider threat domain.
♻ ☆ Interpreting Large Text-to-Image Diffusion Models with Dictionary Learning CVPR 2025
Sparse autoencoders are a promising new approach for decomposing language model activations for interpretation and control. They have been applied successfully to vision transformer image encoders and to small-scale diffusion models. Inference-Time Decomposition of Activations (ITDA) is a recently proposed variant of dictionary learning that takes the dictionary to be a set of data points from the activation distribution and reconstructs them with gradient pursuit. We apply Sparse Autoencoders (SAEs) and ITDA to a large text-to-image diffusion model, Flux 1, and consider the interpretability of embeddings of both by introducing a visual automated interpretation pipeline. We find that SAEs accurately reconstruct residual stream embeddings and beat MLP neurons on interpretability. We are able to use SAE features to steer image generation through activation addition. We find that ITDA has comparable interpretability to SAEs.
comment: 10 pages, 10 figures, Mechanistic Interpretability for Vision at CVPR 2025
♻ ☆ ZKTorch: Compiling ML Inference to Zero-Knowledge Proofs via Parallel Proof Accumulation
As AI models become ubiquitous in our daily lives, there has been an increasing demand for transparency in ML services. However, the model owner does not want to reveal the weights, as they are considered trade secrets. To solve this problem, researchers have turned to zero-knowledge proofs of ML model inference. These proofs convince the user that the ML model output is correct, without revealing the weights of the model to the user. Past work on these provers can be placed into two categories. The first method compiles the ML model into a low-level circuit, and proves the circuit using a ZK-SNARK. The second method uses custom cryptographic protocols designed only for a specific class of models. Unfortunately, the first method is highly inefficient, making it impractical for the large models used today, and the second method does not generalize well, making it difficult to update in the rapidly changing field of machine learning. To solve this, we propose ZKTorch, an open source end-to-end proving system that compiles ML models into base cryptographic operations called basic blocks, each proved using specialized protocols. ZKTorch is built on top of a novel parallel extension to the Mira accumulation scheme, enabling succinct proofs with minimal accumulation overhead. These contributions allow ZKTorch to achieve at least a $3\times$ reduction in the proof size compared to specialized protocols and up to a $6\times$ speedup in proving time over a general-purpose ZKML framework.
comment: 16 pages, 2 figures
♻ ☆ Extracting memorized pieces of (copyrighted) books from open-weight language models
Plaintiffs and defendants in copyright lawsuits over generative AI often make sweeping, opposing claims about the extent to which large language models (LLMs) have memorized plaintiffs' protected expression. Drawing on adversarial ML and copyright law, we show that these polarized positions dramatically oversimplify the relationship between memorization and copyright. To do so, we leverage a recent probabilistic extraction technique to extract pieces of the Books3 dataset from 17 open-weight LLMs. Through numerous experiments, we show that it's possible to extract substantial parts of at least some books from different LLMs. This is evidence that these LLMs have memorized the extracted text; this memorized content is copied inside the model parameters. But the results are complicated: the extent of memorization varies both by model and by book. With our specific experiments, we find that the largest LLMs don't memorize most books--either in whole or in part. However, we also find that Llama 3.1 70B memorizes some books, like Harry Potter and the Sorcerer's Stone and 1984, almost entirely. In fact, Harry Potter is so memorized that, using a seed prompt consisting of just the first line of chapter 1, we can deterministically generate the entire book near-verbatim. We discuss why our results have significant implications for copyright cases, though not ones that unambiguously favor either side.
Multimedia 8
☆ Multimodal Framework for Explainable Autonomous Driving: Integrating Video, Sensor, and Textual Data for Enhanced Decision-Making and Transparency
Autonomous vehicles (AVs) are poised to redefine transportation by enhancing road safety, minimizing human error, and optimizing traffic efficiency. The success of AVs depends on their ability to interpret complex, dynamic environments through diverse data sources, including video streams, sensor measurements, and contextual textual information. However, seamlessly integrating these multimodal inputs and ensuring transparency in AI-driven decisions remain formidable challenges. This study introduces a novel multimodal framework that synergistically combines video, sensor, and textual data to predict driving actions while generating human-readable explanations, fostering trust and regulatory compliance. By leveraging VideoMAE for spatiotemporal video analysis, a custom sensor fusion module for real-time data processing, and BERT for textual comprehension, our approach achieves robust decision-making and interpretable outputs. Evaluated on the BDD-X (21113 samples) and nuScenes (1000 scenes) datasets, our model reduces training loss from 5.7231 to 0.0187 over five epochs, attaining an action prediction accuracy of 92.5% and a BLEU-4 score of 0.75 for explanation quality, outperforming state-of-the-art methods. Ablation studies confirm the critical role of each modality, while qualitative analyses and human evaluations highlight the model's ability to produce contextually rich, user-friendly explanations. These advancements underscore the transformative potential of multimodal integration and explainability in building safe, transparent, and trustworthy AV systems, paving the way for broader societal adoption of autonomous driving technologies.
☆ The Potential of Olfactory Stimuli in Stress Reduction through Virtual Reality
Immersive virtual reality (VR) is a promising tool for stress reduction and relaxation, traditionally relying on visual and auditory stimuli. This study examines the role of olfactory stimuli in enhancing these effects, using a randomized within-subject design. Thirty participants aged 18-60 experienced VR scenarios simulating a calming seaside environment, with sessions lasting 45 minutes, in two conditions: with and without a "Beach" essential oil scent (Yankee Candle) administered via diffuser. Stress and relaxation were assessed through self-reported surveys and physiological measures, specifically ECG-based heart rate variability (HRV). Results showed no significant difference in self-reported relaxation scores (p=0.371) between conditions, but HRV analysis revealed a significant stress reduction (p=0.002) with olfactory input, with HF increasing 108% from the Math Stress Test to the scented relaxation condition, compared to 44% without scent. Additionally, 71.4% of participants expressed willingness to use olfactory-enhanced VR for relaxation, suggesting practical appeal. These findings indicate that olfactory stimuli may enhance relaxation subconsciously, underscoring the importance of multisensory integration in VR. Future work could explore personalized scents and long-term effects to optimize VR- based interventions for emotional and physical well-being.
comment: Accepted to IEEE Medical Measurements & Applications (MeMeA) 2025
☆ T-GVC: Trajectory-Guided Generative Video Coding at Ultra-Low Bitrates
Recent advances in video generation techniques have given rise to an emerging paradigm of generative video coding, aiming to achieve semantically accurate reconstructions in Ultra-Low Bitrate (ULB) scenarios by leveraging strong generative priors. However, most existing methods are limited by domain specificity (e.g., facial or human videos) or an excessive dependence on high-level text guidance, which often fails to capture motion details and results in unrealistic reconstructions. To address these challenges, we propose a Trajectory-Guided Generative Video Coding framework (dubbed T-GVC). T-GVC employs a semantic-aware sparse motion sampling pipeline to effectively bridge low-level motion tracking with high-level semantic understanding by extracting pixel-wise motion as sparse trajectory points based on their semantic importance, not only significantly reducing the bitrate but also preserving critical temporal semantic information. In addition, by incorporating trajectory-aligned loss constraints into diffusion processes, we introduce a training-free latent space guidance mechanism to ensure physically plausible motion patterns without sacrificing the inherent capabilities of generative models. Experimental results demonstrate that our framework outperforms both traditional codecs and state-of-the-art end-to-end video compression methods under ULB conditions. Furthermore, additional experiments confirm that our approach achieves more precise motion control than existing text-guided methods, paving the way for a novel direction of generative video coding guided by geometric motion modeling.
☆ IML-Spikeformer: Input-aware Multi-Level Spiking Transformer for Speech Processing
Spiking Neural Networks (SNNs), inspired by biological neural mechanisms, represent a promising neuromorphic computing paradigm that offers energy-efficient alternatives to traditional Artificial Neural Networks (ANNs). Despite proven effectiveness, SNN architectures have struggled to achieve competitive performance on large-scale speech processing task. Two key challenges hinder progress: (1) the high computational overhead during training caused by multi-timestep spike firing, and (2) the absence of large-scale SNN architectures tailored to speech processing tasks. To overcome the issues, we introduce Input-aware Multi-Level Spikeformer, i.e. IML-Spikeformer, a spiking Transformer architecture specifically designed for large-scale speech processing. Central to our design is the Input-aware Multi-Level Spike (IMLS) mechanism, which simulate multi-timestep spike firing within a single timestep using an adaptive, input-aware thresholding scheme. IML-Spikeformer further integrates a Reparameterized Spiking Self-Attention (RepSSA) module with a Hierarchical Decay Mask (HDM), forming the HD-RepSSA module. This module enhances the precision of attention maps and enables modeling of multi-scale temporal dependencies in speech signals. Experiments demonstrate that IML-Spikeformer achieves word error rates of 6.0\% on AiShell-1 and 3.4\% on Librispeech-960, comparable to conventional ANN transformers while reducing theoretical inference energy consumption by 4.64$\times$ and 4.32$\times$ respectively. IML-Spikeformer marks an advance of scalable SNN architectures for large-scale speech processing in both task performance and energy efficiency.
comment: Under review of TNNLS
☆ PUMA: Layer-Pruned Language Model for Efficient Unified Multimodal Retrieval with Modality-Adaptive Learning ACM MM 2025
As multimedia content expands, the demand for unified multimodal retrieval (UMR) in real-world applications increases. Recent work leverages multimodal large language models (MLLMs) to tackle this task. However, their large parameter size results in high training costs and low inference efficiency. To address this, we propose PUMA: a Layer-Pruned Language Model for Efficient Unified Multimodal Retrieval with Modality-Adaptive Learning. Our approach improves UMR from both structural and learning perspectives. (1) Structurally, we propose Layer-Pruned Self-Distillation, which prunes MLLMs by keeping only shallow layers while distilling features from dropped deep layers as teacher signals. This reduces parameters and preserves representation capability. (2) On the learning side, we introduce Modality-Adaptive Contrastive Learning Loss (MAC-Loss), which separates in-batch negatives into harder intra-modality and easier inter-modality groups based on the target modality, assigning different temperature strategies to enhance learning efficiency. Experiments show our method significantly reduces resource usage while maintaining strong performance.
comment: Accepted to ACM MM 2025
♻ ☆ U-Sticker: A Large-Scale Multi-Domain User Sticker Dataset for Retrieval and Personalization SIGIR'25
Instant messaging with texts and stickers has become a widely adopted communication medium, enabling efficient expression of user semantics and emotions. With the increased use of stickers conveying information and feelings, sticker retrieval and recommendation has emerged as an important area of research. However, a major limitation in existing literature has been the lack of datasets capturing temporal and user-specific sticker interactions, which has hindered further progress in user modeling and sticker personalization. To address this, we introduce User-Sticker, a dataset that includes temporal and user anonymous ID across conversations. It is the largest publicly available sticker dataset to date, containing 22K unique users, 370K stickers, and 8.3M messages. The raw data was collected from a popular messaging platform from 67 conversations over 720 hours of crawling. All text and image data were carefully vetted for safety and privacy checks and modifications. Spanning 10 domains, the U-Sticker dataset captures rich temporal, multilingual, and cross-domain behaviors not previously available in other datasets. Extensive quantitative and qualitative experiments demonstrate U-Sticker's practical applications in user behavior modeling and personalized recommendation and highlight its potential to further research areas in personalized retrieval and conversational studies. U-Sticker dataset is publicly available.
comment: Accepted at SIGIR'25
♻ ☆ HeLo: Heterogeneous Multi-Modal Fusion with Label Correlation for Emotion Distribution Learning
Multi-modal emotion recognition has garnered increasing attention as it plays a significant role in human-computer interaction (HCI) in recent years. Since different discrete emotions may exist at the same time, compared with single-class emotion recognition, emotion distribution learning (EDL) that identifies a mixture of basic emotions has gradually emerged as a trend. However, existing EDL methods face challenges in mining the heterogeneity among multiple modalities. Besides, rich semantic correlations across arbitrary basic emotions are not fully exploited. In this paper, we propose a multi-modal emotion distribution learning framework, named HeLo, aimed at fully exploring the heterogeneity and complementary information in multi-modal emotional data and label correlation within mixed basic emotions. Specifically, we first adopt cross-attention to effectively fuse the physiological data. Then, an optimal transport (OT)-based heterogeneity mining module is devised to mine the interaction and heterogeneity between the physiological and behavioral representations. To facilitate label correlation learning, we introduce a learnable label embedding optimized by correlation matrix alignment. Finally, the learnable label embeddings and label correlation matrices are integrated with the multi-modal representations through a novel label correlation-driven cross-attention mechanism for accurate emotion distribution learning. Experimental results on two publicly available datasets demonstrate the superiority of our proposed method in emotion distribution learning.
♻ ☆ Using AI to Summarize US Presidential Campaign TV Advertisement Videos, 1952-2012
This paper introduces the largest and most comprehensive dataset of US presidential campaign television advertisements, available in digital format. The dataset also includes machine-searchable transcripts and high-quality summaries designed to facilitate a variety of academic research. To date, there has been great interest in collecting and analyzing US presidential campaign advertisements, but the need for manual procurement and annotation led many to rely on smaller subsets. We design a large-scale parallelized, AI-based analysis pipeline that automates the laborious process of preparing, transcribing, and summarizing videos. We then apply this methodology to the 9,707 presidential ads from the Julian P. Kanter Political Commercial Archive. We conduct extensive human evaluations to show that these transcripts and summaries match the quality of manually generated alternatives. We illustrate the value of this data by including an application that tracks the genesis and evolution of current focal issue areas over seven decades of presidential elections. Our analysis pipeline and codebase also show how to use LLM-based tools to obtain high-quality summaries for other video datasets.
comment: 17 pages, 7 tables, 4 figures, and linked datasets
Computation and Language 53
☆ Multi-Agent Retrieval-Augmented Framework for Evidence-Based Counterspeech Against Health Misinformation
Large language models (LLMs) incorporated with Retrieval-Augmented Generation (RAG) have demonstrated powerful capabilities in generating counterspeech against misinformation. However, current studies rely on limited evidence and offer less control over final outputs. To address these challenges, we propose a Multi-agent Retrieval-Augmented Framework to generate counterspeech against health misinformation, incorporating multiple LLMs to optimize knowledge retrieval, evidence enhancement, and response refinement. Our approach integrates both static and dynamic evidence, ensuring that the generated counterspeech is relevant, well-grounded, and up-to-date. Our method outperforms baseline approaches in politeness, relevance, informativeness, and factual accuracy, demonstrating its effectiveness in generating high-quality counterspeech. To further validate our approach, we conduct ablation studies to verify the necessity of each component in our framework. Furthermore, human evaluations reveal that refinement significantly enhances counterspeech quality and obtains human preference.
☆ ViDove: A Translation Agent System with Multimodal Context and Memory-Augmented Reasoning
LLM-based translation agents have achieved highly human-like translation results and are capable of handling longer and more complex contexts with greater efficiency. However, they are typically limited to text-only inputs. In this paper, we introduce ViDove, a translation agent system designed for multimodal input. Inspired by the workflow of human translators, ViDove leverages visual and contextual background information to enhance the translation process. Additionally, we integrate a multimodal memory system and long-short term memory modules enriched with domain-specific knowledge, enabling the agent to perform more accurately and adaptively in real-world scenarios. As a result, ViDove achieves significantly higher translation quality in both subtitle generation and general translation tasks, with a 28% improvement in BLEU scores and a 15% improvement in SubER compared to previous state-of-the-art baselines. Moreover, we introduce DoveBench, a new benchmark for long-form automatic video subtitling and translation, featuring 17 hours of high-quality, human-annotated data. Our code is available here: https://github.com/pigeonai-org/ViDove
☆ The Impact of Background Speech on Interruption Detection in Collaborative Groups
Interruption plays a crucial role in collaborative learning, shaping group interactions and influencing knowledge construction. AI-driven support can assist teachers in monitoring these interactions. However, most previous work on interruption detection and interpretation has been conducted in single-conversation environments with relatively clean audio. AI agents deployed in classrooms for collaborative learning within small groups will need to contend with multiple concurrent conversations -- in this context, overlapping speech will be ubiquitous, and interruptions will need to be identified in other ways. In this work, we analyze interruption detection in single-conversation and multi-group dialogue settings. We then create a state-of-the-art method for interruption identification that is robust to overlapping speech, and thus could be deployed in classrooms. Further, our work highlights meaningful linguistic and prosodic information about how interruptions manifest in collaborative group interactions. Our investigation also paves the way for future works to account for the influence of overlapping speech from multiple groups when tracking group dialog.
comment: Long Paper AIED 2025
☆ LinguaMark: Do Multimodal Models Speak Fairly? A Benchmark-Based Evaluation
Large Multimodal Models (LMMs) are typically trained on vast corpora of image-text data but are often limited in linguistic coverage, leading to biased and unfair outputs across languages. While prior work has explored multimodal evaluation, less emphasis has been placed on assessing multilingual capabilities. In this work, we introduce LinguaMark, a benchmark designed to evaluate state-of-the-art LMMs on a multilingual Visual Question Answering (VQA) task. Our dataset comprises 6,875 image-text pairs spanning 11 languages and five social attributes. We evaluate models using three key metrics: Bias, Answer Relevancy, and Faithfulness. Our findings reveal that closed-source models generally achieve the highest overall performance. Both closed-source (GPT-4o and Gemini2.5) and open-source models (Gemma3, Qwen2.5) perform competitively across social attributes, and Qwen2.5 demonstrates strong generalization across multiple languages. We release our benchmark and evaluation code to encourage reproducibility and further research.
comment: Accepted at ASONAM'25
☆ Open Source Planning & Control System with Language Agents for Autonomous Scientific Discovery ICML 2025
We present a multi-agent system for automation of scientific research tasks, cmbagent. The system is formed by about 30 Large Language Model (LLM) agents and implements a Planning & Control strategy to orchestrate the agentic workflow, with no human-in-the-loop at any point. Each agent specializes in a different task (performing retrieval on scientific papers and codebases, writing code, interpreting results, critiquing the output of other agents) and the system is able to execute code locally. We successfully apply cmbagent to carry out a PhD level cosmology task (the measurement of cosmological parameters using supernova data) and evaluate its performance on two benchmark sets, finding superior performance over state-of-the-art LLMs. The source code is available on GitHub, demonstration videos are also available, and the system is deployed on HuggingFace and will be available on the cloud.
comment: Accepted contribution to the ICML 2025 Workshop on Machine Learning for Astrophysics. Code: https://github.com/CMBAgents/cmbagent; Videos: https://www.youtube.com/@cmbagent; HuggingFace: https://huggingface.co/spaces/astropilot-ai/cmbagent; Cloud: https://cmbagent.cloud
☆ A Language-Driven Framework for Improving Personalized Recommendations: Merging LLMs with Traditional Algorithms
Traditional recommendation algorithms are not designed to provide personalized recommendations based on user preferences provided through text, e.g., "I enjoy light-hearted comedies with a lot of humor". Large Language Models (LLMs) have emerged as one of the most promising tools for natural language processing in recent years. This research proposes a novel framework that mimics how a close friend would recommend items based on their knowledge of an individual's tastes. We leverage LLMs to enhance movie recommendation systems by refining traditional algorithm outputs and integrating them with language-based user preference inputs. We employ Singular Value Decomposition (SVD) or SVD++ algorithms to generate initial movie recommendations, implemented using the Surprise Python library and trained on the MovieLens-Latest-Small dataset. We compare the performance of the base algorithms with our LLM-enhanced versions using leave-one-out validation hit rates and cumulative hit rates. Additionally, to compare the performance of our framework against the current state-of-the-art recommendation systems, we use rating and ranking metrics with an item-based stratified 0.75 train, 0.25 test split. Our framework can generate preference profiles automatically based on users' favorite movies or allow manual preference specification for more personalized results. Using an automated approach, our framework overwhelmingly surpassed SVD and SVD++ on every evaluation metric used (e.g., improvements of up to ~6x in cumulative hit rate, ~3.7x in NDCG, etc.), albeit at the cost of a slight increase in computational overhead.
☆ Medical Red Teaming Protocol of Language Models: On the Importance of User Perspectives in Healthcare Settings
As the performance of large language models (LLMs) continues to advance, their adoption is expanding across a wide range of domains, including the medical field. The integration of LLMs into medical applications raises critical safety concerns, particularly due to their use by users with diverse roles, e.g. patients and clinicians, and the potential for model's outputs to directly affect human health. Despite the domain-specific capabilities of medical LLMs, prior safety evaluations have largely focused only on general safety benchmarks. In this paper, we introduce a safety evaluation protocol tailored to the medical domain in both patient user and clinician user perspectives, alongside general safety assessments and quantitatively analyze the safety of medical LLMs. We bridge a gap in the literature by building the PatientSafetyBench containing 466 samples over 5 critical categories to measure safety from the perspective of the patient. We apply our red-teaming protocols on the MediPhi model collection as a case study. To our knowledge, this is the first work to define safety evaluation criteria for medical LLMs through targeted red-teaming taking three different points of view - patient, clinician, and general user - establishing a foundation for safer deployment in medical domains.
☆ An Information-Theoretic Perspective on Multi-LLM Uncertainty Estimation
Large language models (LLMs) often behave inconsistently across inputs, indicating uncertainty and motivating the need for its quantification in high-stakes settings. Prior work on calibration and uncertainty quantification often focuses on individual models, overlooking the potential of model diversity. We hypothesize that LLMs make complementary predictions due to differences in training and the Zipfian nature of language, and that aggregating their outputs leads to more reliable uncertainty estimates. To leverage this, we propose MUSE (Multi-LLM Uncertainty via Subset Ensembles), a simple information-theoretic method that uses Jensen-Shannon Divergence to identify and aggregate well-calibrated subsets of LLMs. Experiments on binary prediction tasks demonstrate improved calibration and predictive performance compared to single-model and naive ensemble baselines.
comment: Under review
☆ SynthTextEval: Synthetic Text Data Generation and Evaluation for High-Stakes Domains
We present SynthTextEval, a toolkit for conducting comprehensive evaluations of synthetic text. The fluency of large language model (LLM) outputs has made synthetic text potentially viable for numerous applications, such as reducing the risks of privacy violations in the development and deployment of AI systems in high-stakes domains. Realizing this potential, however, requires principled consistent evaluations of synthetic data across multiple dimensions: its utility in downstream systems, the fairness of these systems, the risk of privacy leakage, general distributional differences from the source text, and qualitative feedback from domain experts. SynthTextEval allows users to conduct evaluations along all of these dimensions over synthetic data that they upload or generate using the toolkit's generation module. While our toolkit can be run over any data, we highlight its functionality and effectiveness over datasets from two high-stakes domains: healthcare and law. By consolidating and standardizing evaluation metrics, we aim to improve the viability of synthetic text, and in-turn, privacy-preservation in AI development.
Prompt Perturbations Reveal Human-Like Biases in LLM Survey Responses
Large Language Models (LLMs) are increasingly used as proxies for human subjects in social science surveys, but their reliability and susceptibility to known response biases are poorly understood. This paper investigates the response robustness of LLMs in normative survey contexts -- we test nine diverse LLMs on questions from the World Values Survey (WVS), applying a comprehensive set of 11 perturbations to both question phrasing and answer option structure, resulting in over 167,000 simulated interviews. In doing so, we not only reveal LLMs' vulnerabilities to perturbations but also reveal that all tested models exhibit a consistent \textit{recency bias} varying in intensity, disproportionately favoring the last-presented answer option. While larger models are generally more robust, all models remain sensitive to semantic variations like paraphrasing and to combined perturbations. By applying a set of perturbations, we reveal that LLMs partially align with survey response biases identified in humans. This underscores the critical importance of prompt design and robustness testing when using LLMs to generate synthetic survey data.
comment: 18 pages, 17 figures
☆ Planted in Pretraining, Swayed by Finetuning: A Case Study on the Origins of Cognitive Biases in LLMs
Large language models (LLMs) exhibit cognitive biases -- systematic tendencies of irrational decision-making, similar to those seen in humans. Prior work has found that these biases vary across models and can be amplified by instruction tuning. However, it remains unclear if these differences in biases stem from pretraining, finetuning, or even random noise due to training stochasticity. We propose a two-step causal experimental approach to disentangle these factors. First, we finetune models multiple times using different random seeds to study how training randomness affects over $30$ cognitive biases. Second, we introduce \emph{cross-tuning} -- swapping instruction datasets between models to isolate bias sources. This swap uses datasets that led to different bias patterns, directly testing whether biases are dataset-dependent. Our findings reveal that while training randomness introduces some variability, biases are mainly shaped by pretraining: models with the same pretrained backbone exhibit more similar bias patterns than those sharing only finetuning data. These insights suggest that understanding biases in finetuned models requires considering their pretraining origins beyond finetuning effects. This perspective can guide future efforts to develop principled strategies for evaluating and mitigating bias in LLMs.
comment: CoLM 2025
☆ UniConv: Unifying Retrieval and Response Generation for Large Language Models in Conversations ACL 2025
The rapid advancement of conversational search systems revolutionizes how information is accessed by enabling the multi-turn interaction between the user and the system. Existing conversational search systems are usually built with two different models. This separation restricts the system from leveraging the intrinsic knowledge of the models simultaneously, which cannot ensure the effectiveness of retrieval benefiting the generation. The existing studies for developing unified models cannot fully address the aspects of understanding conversational context, managing retrieval independently, and generating responses. In this paper, we explore how to unify dense retrieval and response generation for large language models in conversation. We conduct joint fine-tuning with different objectives and design two mechanisms to reduce the inconsistency risks while mitigating data discrepancy. The evaluations on five conversational search datasets demonstrate that our unified model can mutually improve both tasks and outperform the existing baselines.
comment: Accepted by ACL 2025 (main)
☆ FlexOlmo: Open Language Models for Flexible Data Use
We introduce FlexOlmo, a new class of language models (LMs) that supports (1) distributed training without data sharing, where different model parameters are independently trained on closed datasets, and (2) data-flexible inference, where these parameters along with their associated data can be flexibly included or excluded from model inferences with no further training. FlexOlmo employs a mixture-of-experts (MoE) architecture where each expert is trained independently on closed datasets and later integrated through a new domain-informed routing without any joint training. FlexOlmo is trained on FlexMix, a corpus we curate comprising publicly available datasets alongside seven domain-specific sets, representing realistic approximations of closed sets. We evaluate models with up to 37 billion parameters (20 billion active) on 31 diverse downstream tasks. We show that a general expert trained on public data can be effectively combined with independently trained experts from other data owners, leading to an average 41% relative improvement while allowing users to opt out of certain data based on data licensing or permission requirements. Our approach also outperforms prior model merging methods by 10.1% on average and surpasses the standard MoE trained without data restrictions using the same training FLOPs. Altogether, this research presents a solution for both data owners and researchers in regulated industries with sensitive or protected data. FlexOlmo enables benefiting from closed data while respecting data owners' preferences by keeping their data local and supporting fine-grained control of data access during inference.
☆ Learning Deliberately, Acting Intuitively: Unlocking Test-Time Reasoning in Multimodal LLMs
Reasoning is a key capability for large language models (LLMs), particularly when applied to complex tasks such as mathematical problem solving. However, multimodal reasoning research still requires further exploration of modality alignment and training costs. Many of these approaches rely on additional data annotation and relevant rule-based rewards to enhance the understanding and reasoning ability, which significantly increases training costs and limits scalability. To address these challenges, we propose the Deliberate-to-Intuitive reasoning framework (D2I) that improves the understanding and reasoning ability of multimodal LLMs (MLLMs) without extra annotations and complex rewards. Specifically, our method sets deliberate reasoning strategies to enhance modality alignment only through the rule-based format reward during training. While evaluating, the reasoning style shifts to intuitive, which removes deliberate reasoning strategies during training and implicitly reflects the model's acquired abilities in the response. D2I outperforms baselines across both in-domain and out-of-domain benchmarks. Our findings highlight the role of format reward in fostering transferable reasoning skills in MLLMs, and inspire directions for decoupling training-time reasoning depth from test-time response flexibility.
comment: Work in progress
☆ FRaN-X: FRaming and Narratives-eXplorer EMNLP 2025
We present FRaN-X, a Framing and Narratives Explorer that automatically detects entity mentions and classifies their narrative roles directly from raw text. FRaN-X comprises a two-stage system that combines sequence labeling with fine-grained role classification to reveal how entities are portrayed as protagonists, antagonists, or innocents, using a unique taxonomy of 22 fine-grained roles nested under these three main categories. The system supports five languages (Bulgarian, English, Hindi, Russian, and Portuguese) and two domains (the Russia-Ukraine Conflict and Climate Change). It provides an interactive web interface for media analysts to explore and compare framing across different sources, tackling the challenge of automatically detecting and labeling how entities are framed. Our system allows end users to focus on a single article as well as analyze up to four articles simultaneously. We provide aggregate level analysis including an intuitive graph visualization that highlights the narrative a group of articles are pushing. Our system includes a search feature for users to look up entities of interest, along with a timeline view that allows analysts to track an entity's role transitions across different contexts within the article. The FRaN-X system and the trained models are licensed under an MIT License. FRaN-X is publicly accessible at https://fran-x.streamlit.app/ and a video demonstration is available at https://youtu.be/VZVi-1B6yYk.
comment: 19 pages, 13 figures, submitted to EMNLP 2025 - Demo Track
☆ Scaling Towards the Information Boundary of Instruction Set: InfinityInstruct-Subject Technical Report
Instruction tuning has become a foundation for unlocking the capabilities of large-scale pretrained models and improving their performance on complex tasks. Thus, the construction of high-quality instruction datasets is crucial for enhancing model performance and generalizability. Although current instruction datasets have reached tens of millions of samples, models finetuned on them may still struggle with complex instruction following and tasks in rare domains. This is primarily due to limited expansion in both ``coverage'' (coverage of task types and knowledge areas) and ``depth'' (instruction complexity) of the instruction set. To address this issue, we propose a systematic instruction data construction framework, which integrates a hierarchical labeling system, an informative seed selection algorithm, an evolutionary data synthesis process, and a model deficiency diagnosis with targeted data generation. These components form an iterative closed-loop to continuously enhance the coverage and depth of instruction data. Based on this framework, we construct InfinityInstruct-Subject, a high-quality dataset containing ~1.5 million instructions. Experiments on multiple foundation models and benchmark tasks demonstrate its effectiveness in improving instruction-following capabilities. Further analyses suggest that InfinityInstruct-Subject shows enlarged coverage and depth compared to comparable synthesized instruction datasets. Our work lays a theoretical and practical foundation for the efficient, continuous evolution of instruction datasets, moving from data quantity expansion to qualitative improvement.
☆ Investigating the Robustness of Retrieval-Augmented Generation at the Query Level ACL 2025
Large language models (LLMs) are very costly and inefficient to update with new information. To address this limitation, retrieval-augmented generation (RAG) has been proposed as a solution that dynamically incorporates external knowledge during inference, improving factual consistency and reducing hallucinations. Despite its promise, RAG systems face practical challenges-most notably, a strong dependence on the quality of the input query for accurate retrieval. In this paper, we investigate the sensitivity of different components in the RAG pipeline to various types of query perturbations. Our analysis reveals that the performance of commonly used retrievers can degrade significantly even under minor query variations. We study each module in isolation as well as their combined effect in an end-to-end question answering setting, using both general-domain and domain-specific datasets. Additionally, we propose an evaluation framework to systematically assess the query-level robustness of RAG pipelines and offer actionable recommendations for practitioners based on the results of more than 1092 experiments we performed.
comment: Accepted to Generation, Evaluation & Metrics (GEM) Workshop at ACL 2025
☆ Exploring LLMs for Predicting Tutor Strategy and Student Outcomes in Dialogues
Tutoring dialogues have gained significant attention in recent years, given the prominence of online learning and the emerging tutoring abilities of artificial intelligence (AI) agents powered by large language models (LLMs). Recent studies have shown that the strategies used by tutors can have significant effects on student outcomes, necessitating methods to predict how tutors will behave and how their actions impact students. However, few works have studied predicting tutor strategy in dialogues. Therefore, in this work we investigate the ability of modern LLMs, particularly Llama 3 and GPT-4o, to predict both future tutor moves and student outcomes in dialogues, using two math tutoring dialogue datasets. We find that even state-of-the-art LLMs struggle to predict future tutor strategy while tutor strategy is highly indicative of student outcomes, outlining a need for more powerful methods to approach this task.
comment: Published in BEA 2025: 20th Workshop on Innovative Use of NLP for Building Educational Applications
☆ MultiJustice: A Chinese Dataset for Multi-Party, Multi-Charge Legal Prediction NLPCC 2025
Legal judgment prediction offers a compelling method to aid legal practitioners and researchers. However, the research question remains relatively under-explored: Should multiple defendants and charges be treated separately in LJP? To address this, we introduce a new dataset namely multi-person multi-charge prediction (MPMCP), and seek the answer by evaluating the performance of several prevailing legal large language models (LLMs) on four practical legal judgment scenarios: (S1) single defendant with a single charge, (S2) single defendant with multiple charges, (S3) multiple defendants with a single charge, and (S4) multiple defendants with multiple charges. We evaluate the dataset across two LJP tasks, i.e., charge prediction and penalty term prediction. We have conducted extensive experiments and found that the scenario involving multiple defendants and multiple charges (S4) poses the greatest challenges, followed by S2, S3, and S1. The impact varies significantly depending on the model. For example, in S4 compared to S1, InternLM2 achieves approximately 4.5% lower F1-score and 2.8% higher LogD, while Lawformer demonstrates around 19.7% lower F1-score and 19.0% higher LogD. Our dataset and code are available at https://github.com/lololo-xiao/MultiJustice-MPMCP.
comment: Accepted by NLPCC 2025
☆ MIND: A Multi-agent Framework for Zero-shot Harmful Meme Detection ACL 2025
The rapid expansion of memes on social media has highlighted the urgent need for effective approaches to detect harmful content. However, traditional data-driven approaches struggle to detect new memes due to their evolving nature and the lack of up-to-date annotated data. To address this issue, we propose MIND, a multi-agent framework for zero-shot harmful meme detection that does not rely on annotated data. MIND implements three key strategies: 1) We retrieve similar memes from an unannotated reference set to provide contextual information. 2) We propose a bi-directional insight derivation mechanism to extract a comprehensive understanding of similar memes. 3) We then employ a multi-agent debate mechanism to ensure robust decision-making through reasoned arbitration. Extensive experiments on three meme datasets demonstrate that our proposed framework not only outperforms existing zero-shot approaches but also shows strong generalization across different model architectures and parameter scales, providing a scalable solution for harmful meme detection. The code is available at https://github.com/destroy-lonely/MIND.
comment: ACL 2025
☆ VisualTrap: A Stealthy Backdoor Attack on GUI Agents via Visual Grounding Manipulation
Graphical User Interface (GUI) agents powered by Large Vision-Language Models (LVLMs) have emerged as a revolutionary approach to automating human-machine interactions, capable of autonomously operating personal devices (e.g., mobile phones) or applications within the device to perform complex real-world tasks in a human-like manner. However, their close integration with personal devices raises significant security concerns, with many threats, including backdoor attacks, remaining largely unexplored. This work reveals that the visual grounding of GUI agent-mapping textual plans to GUI elements-can introduce vulnerabilities, enabling new types of backdoor attacks. With backdoor attack targeting visual grounding, the agent's behavior can be compromised even when given correct task-solving plans. To validate this vulnerability, we propose VisualTrap, a method that can hijack the grounding by misleading the agent to locate textual plans to trigger locations instead of the intended targets. VisualTrap uses the common method of injecting poisoned data for attacks, and does so during the pre-training of visual grounding to ensure practical feasibility of attacking. Empirical results show that VisualTrap can effectively hijack visual grounding with as little as 5% poisoned data and highly stealthy visual triggers (invisible to the human eye); and the attack can be generalized to downstream tasks, even after clean fine-tuning. Moreover, the injected trigger can remain effective across different GUI environments, e.g., being trained on mobile/web and generalizing to desktop environments. These findings underscore the urgent need for further research on backdoor attack risks in GUI agents.
☆ SCoRE: Streamlined Corpus-based Relation Extraction using Multi-Label Contrastive Learning and Bayesian kNN
The growing demand for efficient knowledge graph (KG) enrichment leveraging external corpora has intensified interest in relation extraction (RE), particularly under low-supervision settings. To address the need for adaptable and noise-resilient RE solutions that integrate seamlessly with pre-trained large language models (PLMs), we introduce SCoRE, a modular and cost-effective sentence-level RE system. SCoRE enables easy PLM switching, requires no finetuning, and adapts smoothly to diverse corpora and KGs. By combining supervised contrastive learning with a Bayesian k-Nearest Neighbors (kNN) classifier for multi-label classification, it delivers robust performance despite the noisy annotations of distantly supervised corpora. To improve RE evaluation, we propose two novel metrics: Correlation Structure Distance (CSD), measuring the alignment between learned relational patterns and KG structures, and Precision at R (P@R), assessing utility as a recommender system. We also release Wiki20d, a benchmark dataset replicating real-world RE conditions where only KG-derived annotations are available. Experiments on five benchmarks show that SCoRE matches or surpasses state-of-the-art methods while significantly reducing energy consumption. Further analyses reveal that increasing model complexity, as seen in prior work, degrades performance, highlighting the advantages of SCoRE's minimal design. Combining efficiency, modularity, and scalability, SCoRE stands as an optimal choice for real-world RE applications.
☆ Developing and Maintaining an Open-Source Repository of AI Evaluations: Challenges and Insights
AI evaluations have become critical tools for assessing large language model capabilities and safety. This paper presents practical insights from eight months of maintaining $inspect\_evals$, an open-source repository of 70+ community-contributed AI evaluations. We identify key challenges in implementing and maintaining AI evaluations and develop solutions including: (1) a structured cohort management framework for scaling community contributions, (2) statistical methodologies for optimal resampling and cross-model comparison with uncertainty quantification, and (3) systematic quality control processes for reproducibility. Our analysis reveals that AI evaluation requires specialized infrastructure, statistical rigor, and community coordination beyond traditional software development practices.
☆ Adaptive Termination for Multi-round Parallel Reasoning: An Universal Semantic Entropy-Guided Framework
Recent advances in large language models (LLMs) have accelerated progress toward artificial general intelligence, with inference-time scaling emerging as a key technique. Contemporary approaches leverage either sequential reasoning (iteratively extending chains of thought) or parallel reasoning (generating multiple solutions simultaneously) to scale inference. However, both paradigms face fundamental limitations: sequential scaling typically relies on arbitrary token budgets for termination, leading to inefficiency or premature cutoff; while parallel scaling often lacks coordination among parallel branches and requires intrusive fine-tuning to perform effectively. In light of these challenges, we aim to design a flexible test-time collaborative inference framework that exploits the complementary strengths of both sequential and parallel reasoning paradigms. Towards this goal, the core challenge lies in developing an efficient and accurate intrinsic quality metric to assess model responses during collaborative inference, enabling dynamic control and early termination of the reasoning trace. To address this challenge, we introduce semantic entropy (SE), which quantifies the semantic diversity of parallel model responses and serves as a robust indicator of reasoning quality due to its strong negative correlation with accuracy...
comment: 13 pages, 5 fiures
☆ Text to model via SysML: Automated generation of dynamical system computational models from unstructured natural language text via enhanced System Modeling Language diagrams
This paper contributes to speeding up the design and deployment of engineering dynamical systems by proposing a strategy for exploiting domain and expert knowledge for the automated generation of dynamical system computational model starting from a corpus of document relevant to the dynamical system of interest and an input document describing the specific system. This strategy is implemented in five steps and, crucially, it uses system modeling language diagrams (SysML) to extract accurate information about the dependencies, attributes, and operations of components. Natural Language Processing (NLP) strategies and Large Language Models (LLMs) are employed in specific tasks to improve intermediate outputs of the SySML diagrams automated generation, such as: list of key nouns; list of extracted relationships; list of key phrases and key relationships; block attribute values; block relationships; and BDD diagram generation. The applicability of automated SysML diagram generation is illustrated with different case studies. The computational models of complex dynamical systems from SysML diagrams are then obtained via code generation and computational model generation steps. In the code generation step, NLP strategies are used for summarization, while LLMs are used for validation only. The proposed approach is not limited to a specific system, domain, or computational software. The applicability of the proposed approach is shown via an end-to-end example from text to model of a simple pendulum, showing improved performance compared to results yielded by LLMs only.
☆ Checklist Engineering Empowers Multilingual LLM Judges
Automated text evaluation has long been a central issue in Natural Language Processing (NLP). Recently, the field has shifted toward using Large Language Models (LLMs) as evaluators-a trend known as the LLM-as-a-Judge paradigm. While promising and easily adaptable across tasks, this approach has seen limited exploration in multilingual contexts. Existing multilingual studies often rely on proprietary models or require extensive training data for fine-tuning, raising concerns about cost, time, and efficiency. In this paper, we propose Checklist Engineering based LLM-as-a-Judge (CE-Judge), a training-free framework that uses checklist intuition for multilingual evaluation with an open-source model. Experiments across multiple languages and three benchmark datasets, under both pointwise and pairwise settings, show that our method generally surpasses the baselines and performs on par with the GPT-4o model.
☆ KAConvText: Novel Approach to Burmese Sentence Classification using Kolmogorov-Arnold Convolution
This paper presents the first application of Kolmogorov-Arnold Convolution for Text (KAConvText) in sentence classification, addressing three tasks: imbalanced binary hate speech detection, balanced multiclass news classification, and imbalanced multiclass ethnic language identification. We investigate various embedding configurations, comparing random to fastText embeddings in both static and fine-tuned settings, with embedding dimensions of 100 and 300 using CBOW and Skip-gram models. Baselines include standard CNNs and CNNs augmented with a Kolmogorov-Arnold Network (CNN-KAN). In addition, we investigated KAConvText with different classification heads - MLP and KAN, where using KAN head supports enhanced interpretability. Results show that KAConvText-MLP with fine-tuned fastText embeddings achieves the best performance of 91.23% accuracy (F1-score = 0.9109) for hate speech detection, 92.66% accuracy (F1-score = 0.9267) for news classification, and 99.82% accuracy (F1-score = 0.9982) for language identification.
comment: 10 pages, 3 figures, 4 tables
☆ Robust Multimodal Large Language Models Against Modality Conflict ICML 2025
Despite the impressive capabilities of multimodal large language models (MLLMs) in vision-language tasks, they are prone to hallucinations in real-world scenarios. This paper investigates the hallucination phenomenon in MLLMs from the perspective of modality conflict. Unlike existing works focusing on the conflicts between model responses and inputs, we study the inherent conflicts in inputs from different modalities that place MLLMs in a dilemma and directly lead to hallucinations. We formally define the modality conflict and construct a dataset named Multimodal Modality Conflict (MMMC) to simulate this phenomenon in vision-language tasks. Three methods based on prompt engineering, supervised fine-tuning, and reinforcement learning are proposed to alleviate the hallucination caused by modality conflict. Extensive experiments are conducted on the MMMC dataset to analyze the merits and demerits of these methods. Our results show that the reinforcement learning method achieves the best performance in mitigating the hallucination under modality conflict, while the supervised fine-tuning method shows promising and stable performance. Our work sheds light on the unnoticed modality conflict that leads to hallucinations and provides more insights into the robustness of MLLMs.
comment: ICML 2025
♻ ☆ Taming the Tri-Space Tension: ARC-Guided Hallucination Modeling and Control for Text-to-Image Generation
Despite remarkable progress in image quality and prompt fidelity, text-to-image (T2I) diffusion models continue to exhibit persistent "hallucinations", where generated content subtly or significantly diverges from the intended prompt semantics. While often regarded as unpredictable artifacts, we argue that these failures reflect deeper, structured misalignments within the generative process. In this work, we propose a cognitively inspired perspective that reinterprets hallucinations as trajectory drift within a latent alignment space. Empirical observations reveal that generation unfolds within a multiaxial cognitive tension field, where the model must continuously negotiate competing demands across three key critical axes: semantic coherence, structural alignment, and knowledge grounding. We then formalize this three-axis space as the \textbf{Hallucination Tri-Space} and introduce the Alignment Risk Code (ARC): a dynamic vector representation that quantifies real-time alignment tension during generation. The magnitude of ARC captures overall misalignment, its direction identifies the dominant failure axis, and its imbalance reflects tension asymmetry. Based on this formulation, we develop the TensionModulator (TM-ARC): a lightweight controller that operates entirely in latent space. TM-ARC monitors ARC signals and applies targeted, axis-specific interventions during the sampling process. Extensive experiments on standard T2I benchmarks demonstrate that our approach significantly reduces hallucination without compromising image quality or diversity. This framework offers a unified and interpretable approach for understanding and mitigating generative failures in diffusion-based T2I systems.
comment: We withdraw this paper due to significant visualization errors in Figure 3 and 5 that affect the correctness of our core modeling claims and may cause misinterpretation. These figures misrepresent ARC dynamics and trajectory control
♻ ☆ Rankers, Judges, and Assistants: Towards Understanding the Interplay of LLMs in Information Retrieval Evaluation SIGIR
Large language models (LLMs) are increasingly integral to information retrieval (IR), powering ranking, evaluation, and AI-assisted content creation. This widespread adoption necessitates a critical examination of potential biases arising from the interplay between these LLM-based components. This paper synthesizes existing research and presents novel experiment designs that explore how LLM-based rankers and assistants influence LLM-based judges. We provide the first empirical evidence of LLM judges exhibiting significant bias towards LLM-based rankers. Furthermore, we observe limitations in LLM judges' ability to discern subtle system performance differences. Contrary to some previous findings, our preliminary study does not find evidence of bias against AI-generated content. These results highlight the need for a more holistic view of the LLM-driven information ecosystem. To this end, we offer initial guidelines and a research agenda to ensure the reliable use of LLMs in IR evaluation.
comment: Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '25)
♻ ☆ A Comprehensive Survey of Contamination Detection Methods in Large Language Models
With the rise of Large Language Models (LLMs) in recent years, abundant new opportunities are emerging, but also new challenges, among which contamination is quickly becoming critical. Business applications and fundraising in Artificial Intelligence (AI) have reached a scale at which a few percentage points gained on popular question-answering benchmarks could translate into dozens of millions of dollars, placing high pressure on model integrity. At the same time, it is becoming harder and harder to keep track of the data that LLMs have seen; if not impossible with closed-source models like GPT-4 and Claude-3 not divulging any information on the training set. As a result, contamination becomes a major issue: LLMs' performance may not be reliable anymore, as the high performance may be at least partly due to their previous exposure to the data. This limitation jeopardizes real capability improvement in the field of NLP, yet, there remains a lack of methods on how to efficiently detect contamination. In this paper, we survey all recent work on contamination detection with LLMs, analyzing their methodologies and use cases to shed light on the appropriate usage of contamination detection methods. Our work calls the NLP research community's attention into systematically taking into account contamination bias in LLM evaluation.
comment: Accepted by TMLR in July 2025. 18 pages, 1 figure, 3 tables
♻ ☆ Empowering Healthcare Practitioners with Language Models: Structuring Speech Transcripts in Two Real-World Clinical Applications
Large language models (LLMs) such as GPT-4o and o1 have demonstrated strong performance on clinical natural language processing (NLP) tasks across multiple medical benchmarks. Nonetheless, two high-impact NLP tasks - structured tabular reporting from nurse dictations and medical order extraction from doctor-patient consultations - remain underexplored due to data scarcity and sensitivity, despite active industry efforts. Practical solutions to these real-world clinical tasks can significantly reduce the documentation burden on healthcare providers, allowing greater focus on patient care. In this paper, we investigate these two challenging tasks using private and open-source clinical datasets, evaluating the performance of both open- and closed-weight LLMs, and analyzing their respective strengths and limitations. Furthermore, we propose an agentic pipeline for generating realistic, non-sensitive nurse dictations, enabling structured extraction of clinical observations. To support further research in both areas, we release SYNUR and SIMORD, the first open-source datasets for nurse observation extraction and medical order extraction.
♻ ☆ A Theory of Response Sampling in LLMs: Part Descriptive and Part Prescriptive ACL 2025
Large Language Models (LLMs) are increasingly utilized in autonomous decision-making, where they sample options from vast action spaces. However, the heuristics that guide this sampling process remain under explored. We study this sampling behavior and show that this underlying heuristics resembles that of human decision-making: comprising a descriptive component (reflecting statistical norm) and a prescriptive component (implicit ideal encoded in the LLM) of a concept. We show that this deviation of a sample from the statistical norm towards a prescriptive component consistently appears in concepts across diverse real-world domains like public health, and economic trends. To further illustrate the theory, we demonstrate that concept prototypes in LLMs are affected by prescriptive norms, similar to the concept of normality in humans. Through case studies and comparison with human studies, we illustrate that in real-world applications, the shift of samples toward an ideal value in LLMs' outputs can result in significantly biased decision-making, raising ethical concerns.
comment: ACL 2025 (Oral)
♻ ☆ Emoji Attack: Enhancing Jailbreak Attacks Against Judge LLM Detection
Jailbreaking techniques trick Large Language Models (LLMs) into producing restricted output, posing a potential threat. One line of defense is to use another LLM as a Judge to evaluate the harmfulness of generated text. However, we reveal that these Judge LLMs are vulnerable to token segmentation bias, an issue that arises when delimiters alter the tokenization process, splitting words into smaller sub-tokens. This alters the embeddings of the entire sequence, reducing detection accuracy and allowing harmful content to be misclassified as safe. In this paper, we introduce Emoji Attack, a novel strategy that amplifies existing jailbreak prompts by exploiting token segmentation bias. Our method leverages in-context learning to systematically insert emojis into text before it is evaluated by a Judge LLM, inducing embedding distortions that significantly lower the likelihood of detecting unsafe content. Unlike traditional delimiters, emojis also introduce semantic ambiguity, making them particularly effective in this attack. Through experiments on state-of-the-art Judge LLMs, we demonstrate that Emoji Attack substantially reduces the unsafe prediction rate, bypassing existing safeguards.
♻ ☆ EduCoder: An Open-Source Annotation System for Education Transcript Data
We introduce EduCoder, a domain-specialized tool designed to support utterance-level annotation of educational dialogue. While general-purpose text annotation tools for NLP and qualitative research abound, few address the complexities of coding education dialogue transcripts -- with diverse teacher-student and peer interactions. Common challenges include defining codebooks for complex pedagogical features, supporting both open-ended and categorical coding, and contextualizing utterances with external features, such as the lesson's purpose and the pedagogical value of the instruction. EduCoder is designed to address these challenges by providing a platform for researchers and domain experts to collaboratively define complex codebooks based on observed data. It incorporates both categorical and open-ended annotation types along with contextual materials. Additionally, it offers a side-by-side comparison of multiple annotators' responses, allowing comparison and calibration of annotations with others to improve data reliability. The system is open-source, with a demo video available.
♻ ☆ Derailer-Rerailer: Adaptive Verification for Efficient and Reliable Language Model Reasoning
Large Language Models (LLMs) have shown impressive reasoning capabilities, yet existing prompting methods face a critical trade-off: simple approaches often struggle with complex tasks and reasoning stability, while more sophisticated methods require multiple inferences and substantial computational resources, limiting their practical deployment. To address this challenge, we propose Derailer-Rerailer, a novel framework that adaptively balances reasoning accuracy and computational efficiency. At its core, our framework employs a lightweight Derailer mechanism to assess reasoning stability and selectively triggers an advanced Rerailer verification process only when necessary, thereby optimizing computational resource usage. Extensive evaluation across both open and closed-source models on more than 20 categories of mathematical, symbolic, and commonsense reasoning tasks demonstrates our framework's effectiveness: Derailer-Rerailer achieves significant accuracy improvements (8-11\% across various reasoning tasks) while maintaining 2-3 times better efficiency than existing verification methods, with particularly strong performance in mathematical and symbolic reasoning, offering a practical solution for enhancing LLM reasoning reliability while significantly reducing computational overhead.
♻ ☆ Multi-Attribute Steering of Language Models via Targeted Intervention ACL 2025
Inference-time intervention (ITI) has emerged as a promising method for steering large language model (LLM) behavior in a particular direction (e.g., improving helpfulness) by intervening on token representations without costly updates to the LLM's parameters. However, existing ITI approaches fail to scale to multi-attribute settings with conflicts, such as enhancing helpfulness while also reducing toxicity. To address this, we introduce Multi-Attribute Targeted Steering (MAT-Steer), a novel steering framework designed for selective token-level intervention across multiple attributes. MAT-Steer learns steering vectors using an alignment objective that shifts the model's internal representations of undesirable outputs closer to those of desirable ones while enforcing sparsity and orthogonality among vectors for different attributes, thereby reducing inter-attribute conflicts. We evaluate MAT-Steer in two distinct settings: (i) on question answering (QA) tasks where we balance attributes like truthfulness, bias, and toxicity; (ii) on generative tasks where we simultaneously improve attributes like helpfulness, correctness, and coherence. MAT-Steer outperforms existing ITI and parameter-efficient fine-tuning approaches across both task types (e.g., 3% average accuracy gain across QA tasks and 55.82% win rate against the best ITI baseline).
comment: ACL 2025 camera-ready, code link: https://github.com/duykhuongnguyen/MAT-Steer
♻ ☆ LCFO: Long Context and Long Form Output Dataset and Benchmarking
This paper presents the Long Context and Form Output (LCFO) benchmark, a novel evaluation framework for assessing gradual summarization and summary expansion capabilities across diverse domains. LCFO consists of long input documents (5k words average length), each of which comes with three summaries of different lengths (20%, 10%, and 5% of the input text), as well as approximately 15 questions and answers (QA) related to the input content. Notably, LCFO also provides alignments between specific QA pairs and corresponding summaries in 7 domains. The primary motivation behind providing summaries of different lengths is to establish a controllable framework for generating long texts from shorter inputs, i.e. summary expansion. To establish an evaluation metric framework for summarization and summary expansion, we provide human evaluation scores for human-generated outputs, as well as results from various state-of-the-art large language models (LLMs). GPT-4o-mini achieves best human scores among automatic systems in both summarization and summary expansion tasks (~ +10% and +20%, respectively). It even surpasses human output quality in the case of short summaries (~ +7%). Overall automatic metrics achieve low correlations with human evaluation scores (~ 0.4) but moderate correlation on specific evaluation aspects such as fluency and attribution (~ 0.6).
♻ ☆ LASeR: Learning to Adaptively Select Reward Models with Multi-Armed Bandits
Reward Models (RMs) are crucial to aligning large language models (LLMs), but the degree to which an RM specialized to one task (e.g. writing) generalizes to new tasks (e.g. math) is often not known a priori, often making using only one fixed RM to train LLMs suboptimal. However, optimizing LLMs with multiple RMs simultaneously can incur a prohibitively high computational cost and lead to conflicting signals from different RMs that may degrade performance. To address these challenges, we introduce LASeR (Learning to Adaptively Select Rewards), which frames reward model selection as a multi-armed bandit problem, efficiently and iteratively training LLMs using multiple RMs by selecting the most well-suited RM for each instance. On commonsense and math reasoning tasks, we show that LASeR boosts iterative LLM training, improving the absolute average accuracy of Llama-3-8B over three datasets by 2.67% over an ensemble of RM scores while also showing superior efficiency (e.g., a 2x speedup). Moreover, on WildChat (open-ended instruction-following tasks), LASeR leads to a 72.69% AlpacaEval win rate over the RM score ensemble baseline. Extending to long-context generation, LASeR improves by 2.96 F1 points (avg.) on single-document QA tasks and 2.97 F1 points on few-shot learning over the RM score ensemble baseline with best-of-n sampling.
comment: 28 pages; First two authors contributed equally. Code: https://github.com/duykhuongnguyen/LASeR-MAB
♻ ☆ Low-Rank Adaptation Secretly Imitates Differentially Private SGD
As pre-trained language models grow in size, full fine-tuning their parameters on task adaptation data becomes increasingly impractical. To address this challenge, some methods for low-rank adaptation of language models have been proposed, e.g. LoRA, which incorporates trainable low-rank decomposition matrices into only some parameters of the pre-trained model, called adapters. This approach significantly reduces the number of trainable parameters compared to fine-tuning all parameters or adapters. In this work, we look at low-rank adaptation method from the lens of data privacy. We show theoretically that the low-rank adaptation used in LoRA is equivalent to fine-tuning adapters with noisy batch gradients - just like what DPSGD algorithm does. We also quantify the variance of the injected noise as a decreasing function of adaptation rank. By establishing a Berry-Esseen type bound on the total variation distance between the injected noise distribution and a Gaussian noise distribution with the same variance, we show that the dynamics of low-rank adaptation is very close to when DPSGD is performed w.r.t the adapters. Following our theoretical findings and approved by our experimental results, we show that low-rank adaptation provides robustness to membership inference attacks w.r.t the fine-tuning data.
♻ ☆ TokenShapley: Token Level Context Attribution with Shapley Value
Large language models (LLMs) demonstrate strong capabilities in in-context learning, but verifying the correctness of their generated responses remains a challenge. Prior work has explored attribution at the sentence level, but these methods fall short when users seek attribution for specific keywords within the response, such as numbers, years, or names. To address this limitation, we propose TokenShapley, a novel token-level attribution method that combines Shapley value-based data attribution with KNN-based retrieval techniques inspired by recent advances in KNN-augmented LLMs. By leveraging a precomputed datastore for contextual retrieval and computing Shapley values to quantify token importance, TokenShapley provides a fine-grained data attribution approach. Extensive evaluations on four benchmarks show that TokenShapley outperforms state-of-the-art baselines in token-level attribution, achieving an 11-23% improvement in accuracy.
♻ ☆ Planning Anything with Rigor: General-Purpose Zero-Shot Planning with LLM-based Formalized Programming
While large language models (LLMs) have recently demonstrated strong potential in solving planning problems, there is a trade-off between flexibility and complexity. LLMs, as zero-shot planners themselves, are still not capable of directly generating valid plans for complex planning problems such as multi-constraint or long-horizon tasks. On the other hand, many frameworks aiming to solve complex planning problems often rely on task-specific preparatory efforts, such as task-specific in-context examples and pre-defined critics/verifiers, which limits their cross-task generalization capability. In this paper, we tackle these challenges by observing that the core of many planning problems lies in optimization problems: searching for the optimal solution (best plan) with goals subject to constraints (preconditions and effects of decisions). With LLMs' commonsense, reasoning, and programming capabilities, this opens up the possibilities of a universal LLM-based approach to planning problems. Inspired by this observation, we propose LLMFP, a general-purpose framework that leverages LLMs to capture key information from planning problems and formally formulate and solve them as optimization problems from scratch, with no task-specific examples needed. We apply LLMFP to 9 planning problems, ranging from multi-constraint decision making to multi-step planning problems, and demonstrate that LLMFP achieves on average 83.7% and 86.8% optimal rate across 9 tasks for GPT-4o and Claude 3.5 Sonnet, significantly outperforming the best baseline (direct planning with OpenAI o1-preview) with 37.6% and 40.7% improvements. We also validate components of LLMFP with ablation experiments and analyzed the underlying success and failure reasons. Project page: https://sites.google.com/view/llmfp.
comment: 57 pages, 25 figures, 15 tables
♻ ☆ Neuron-Level Differentiation of Memorization and Generalization in Large Language Models
We investigate how Large Language Models (LLMs) distinguish between memorization and generalization at the neuron level. Through carefully designed tasks, we identify distinct neuron subsets responsible for each behavior. Experiments on both a GPT-2 model trained from scratch and a pretrained LLaMA-3.2 model fine-tuned with LoRA show consistent neuron-level specialization. We further demonstrate that inference-time interventions on these neurons can steer the model's behavior toward memorization or generalization. To assess robustness, we evaluate intra-task and inter-task consistency, confirming that these neuron-behavior associations reflect generalizable patterns rather than dataset-specific artifacts. Our findings reveal modular structure in LLMs and enable controlling memorization and generalization behaviors at inference time.
♻ ☆ Towards Reasoning Era: A Survey of Long Chain-of-Thought for Reasoning Large Language Models
Recent advancements in reasoning with large language models (RLLMs), such as OpenAI-O1 and DeepSeek-R1, have demonstrated their impressive capabilities in complex domains like mathematics and coding. A central factor in their success lies in the application of long chain-of-thought (Long CoT) characteristics, which enhance reasoning abilities and enable the solution of intricate problems. However, despite these developments, a comprehensive survey on Long CoT is still lacking, limiting our understanding of its distinctions from traditional short chain-of-thought (Short CoT) and complicating ongoing debates on issues like "overthinking" and "inference-time scaling." This survey seeks to fill this gap by offering a unified perspective on Long CoT. (1) We first distinguish Long CoT from Short CoT and introduce a novel taxonomy to categorize current reasoning paradigms. (2) Next, we explore the key characteristics of Long CoT: deep reasoning, extensive exploration, and feasible reflection, which enable models to handle more complex tasks and produce more efficient, coherent outcomes compared to the shallower Short CoT. (3) We then investigate key phenomena such as the emergence of Long CoT with these characteristics, including overthinking, and inference-time scaling, offering insights into how these processes manifest in practice. (4) Finally, we identify significant research gaps and highlight promising future directions, including the integration of multi-modal reasoning, efficiency improvements, and enhanced knowledge frameworks. By providing a structured overview, this survey aims to inspire future research and further the development of logical reasoning in artificial intelligence.
comment: Paper are available at https://long-cot.github.io/, and Github are available at https://github.com/LightChen233/Awesome-Long-Chain-of-Thought-Reasoning
♻ ☆ What to Keep and What to Drop: Adaptive Table Filtering Framework
Large language models (LLMs) for table-based reasoning often struggle with large tables due to input length limits. We propose ATF (Adaptive Table Filtering Framework), a modular and question-aware filtering pipeline that prunes uninformative columns and rows using LLM-generated column descriptions, clustering, and sparse-dense alignment scores. ATF integrates seamlessly with existing models (e.g., TAPAS, TAPEX) without retraining. Experiments show that ATF reduces table cells by 70%, boosting performance on out-of-domain TableQA tasks while causing slight performance drops on Table Fact Verification, where full-table context is more critical. These results highlight ATF's ability to adaptively balance informativeness and minimalism across tasks.
comment: 26 pages, 9 figures
♻ ☆ NoLiMa: Long-Context Evaluation Beyond Literal Matching ICML 2025
Recent large language models (LLMs) support long contexts ranging from 128K to 1M tokens. A popular method for evaluating these capabilities is the needle-in-a-haystack (NIAH) test, which involves retrieving a "needle" (relevant information) from a "haystack" (long irrelevant context). Extensions of this approach include increasing distractors, fact chaining, and in-context reasoning. However, in these benchmarks, models can exploit existing literal matches between the needle and haystack to simplify the task. To address this, we introduce NoLiMa, a benchmark extending NIAH with a carefully designed needle set, where questions and needles have minimal lexical overlap, requiring models to infer latent associations to locate the needle within the haystack. We evaluate 13 popular LLMs that claim to support contexts of at least 128K tokens. While they perform well in short contexts (<1K), performance degrades significantly as context length increases. At 32K, for instance, 11 models drop below 50% of their strong short-length baselines. Even GPT-4o, one of the top-performing exceptions, experiences a reduction from an almost-perfect baseline of 99.3% to 69.7%. Our analysis suggests these declines stem from the increased difficulty the attention mechanism faces in longer contexts when literal matches are absent, making it harder to retrieve relevant information. Even models enhanced with reasoning capabilities or CoT prompting struggle to maintain performance in long contexts. We publicly release the dataset and evaluation code at https://github.com/adobe-research/NoLiMa.
comment: Accepted at ICML 2025
♻ ☆ Adaptive Elicitation of Latent Information Using Natural Language ICML 2025
Eliciting information to reduce uncertainty about a latent entity is a critical task in many application domains, e.g., assessing individual student learning outcomes, diagnosing underlying diseases, or learning user preferences. Though natural language is a powerful medium for this purpose, large language models (LLMs) and existing fine-tuning algorithms lack mechanisms for strategically gathering information to refine their own understanding of the latent entity. To harness the generalization power and world knowledge of LLMs in developing effective information-gathering strategies, we propose an adaptive elicitation framework that actively reduces uncertainty on the latent entity. Since probabilistic modeling of an abstract latent entity is difficult, our framework adopts a predictive view of uncertainty, using a meta-learned language model to simulate future observations and enable scalable uncertainty quantification over complex natural language. Through autoregressive forward simulation, our model quantifies how new questions reduce epistemic uncertainty, enabling the development of sophisticated information-gathering strategies to choose the most informative next queries. In experiments on the 20 questions game, dynamic opinion polling, and adaptive student assessment, our method consistently outperforms baselines in identifying critical unknowns and improving downstream predictions, illustrating the promise of strategic information gathering in natural language settings.
comment: ICML 2025
♻ ☆ EMORL: Ensemble Multi-Objective Reinforcement Learning for Efficient and Flexible LLM Fine-Tuning SIGDIAL 2025
Recent advances in reinforcement learning (RL) for large language model (LLM) fine-tuning show promise in addressing multi-objective tasks but still face significant challenges, including competing objective balancing, low training efficiency, poor scalability, and limited explainability. Leveraging ensemble learning principles, we introduce an Ensemble Multi-Objective RL (EMORL) framework that fine-tunes multiple models with individual objectives while optimizing their aggregation after the fine-tuning to improve efficiency and flexibility. Our method is the first to aggregate the hidden states of individual models, incorporating contextual information from multiple objectives. This approach is supported by a hierarchical grid search algorithm that identifies optimal weighted combinations. We evaluate EMORL on counselor reflection generation tasks, using text classification models to score the generations and provide rewards during RL fine-tuning. Through comprehensive experiments on the PAIR and Psych8k datasets, we demonstrate the advantages of EMORL against existing baselines: significantly lower and more stable training consumption ($17,529\pm 1,650$ data points and $6,573\pm 147.43$ seconds), improved scalability and explainability, and comparable performance across multiple objectives.
comment: 14 pages, 9 figures, accepted by the SIGDIAL 2025 conference
♻ ☆ CMQCIC-Bench: A Chinese Benchmark for Evaluating Large Language Models in Medical Quality Control Indicator Calculation ACL
Medical quality control indicators are essential to assess the qualifications of healthcare institutions for medical services. With the impressive performance of large language models (LLMs) like GPT-4 in the medical field, leveraging these technologies for the Medical Quality Control Indicator Calculation (MQCIC) presents a promising approach. In this work, (1) we introduce a real-world task MQCIC and propose an open-source Chinese electronic medical records (EMRs)-based dataset (CMQCIC-Bench) comprising 785 instances and 76 indicators. (2) We propose a semi-automatic method to enhance the rule representation. Then we propose the Clinical Facts-based Inferential Rule (CF-IR) method that disentangles the clinical fact verification and inferential rule reasoning actions. (3) We conduct comprehensive experiments on 20 representative LLMs, covering general and medical models. Our findings reveal that CF-IR outperforms Chain-of-Thought methods in MQCIC tasks. (4) We conduct an error analysis and investigate the capabilities of clinical fact verification and inferential rule reasoning, providing insights to improve performance in the MQCIC further. The dataset and code is available in this repository https://github.com/YuY-2001/C-MQCIC.
comment: 2025 ACL Findings
♻ ☆ Losing our Tail -- Again: On (Un)Natural Selection And Multilingual Large Language Models
Multilingual Large Language Models (LLMs) considerably changed how technologies can influence language. While previous technologies could mediate or assist humans, there is now a tendency to offload the task of writing itself to these technologies, enabling them to change our linguistic ecosystem more directly. While they provide us quick access to information and impressively fluent output, beneath their apparent sophistication lies a subtle, more insidious threat: the gradual decline and loss of linguistic diversity. With this opinion piece, I explore how model collapse, with a particular focus on translation technology, can lead to the loss of linguistic forms, grammatical features, and cultural nuance. Model collapse refers to the eventual consequence of self-consuming training loops, where models reinforce their own biases and lose linguistic diversity. Drawing on recent work in Computer Vision, Natural Language Processing (NLP) and Machine Translation (MT), I argue that the tails of our linguistic distributions are vanishing, and with them, the narratives and identities they carry. This is a call to resist linguistic flattening and to reimagine NLP as a field that encourages, values and protects expressive multilingual lexical and linguistic diversity and creativity.
comment: 12 pages
♻ ☆ Safer or Luckier? LLMs as Safety Evaluators Are Not Robust to Artifacts ACL 2025
Large Language Models (LLMs) are increasingly employed as automated evaluators to assess the safety of generated content, yet their reliability in this role remains uncertain. This study evaluates a diverse set of 11 LLM judge models across critical safety domains, examining three key aspects: self-consistency in repeated judging tasks, alignment with human judgments, and susceptibility to input artifacts such as apologetic or verbose phrasing. Our findings reveal that biases in LLM judges can significantly distort the final verdict on which content source is safer, undermining the validity of comparative evaluations. Notably, apologetic language artifacts alone can skew evaluator preferences by up to 98\%. Contrary to expectations, larger models do not consistently exhibit greater robustness, while smaller models sometimes show higher resistance to specific artifacts. To mitigate LLM evaluator robustness issues, we investigate jury-based evaluations aggregating decisions from multiple models. Although this approach both improves robustness and enhances alignment to human judgements, artifact sensitivity persists even with the best jury configurations. These results highlight the urgent need for diversified, artifact-resistant methodologies to ensure reliable safety assessments.
comment: 9 pages, ACL 2025
♻ ☆ Test-Time Scaling with Reflective Generative Model
We introduce our first reflective generative model MetaStone-S1, which obtains OpenAI o3-mini's performance via the new Reflective Generative Form. The new form focuses on high-quality reasoning trajectory selection and contains two novelties: 1) A unified interface for policy and process reward model: we share the backbone network and use task-specific heads for reasoning trajectory predicting and scoring respectively, introducing only 53M extra parameters for trajectory scoring. 2) Eliminating the reliance on process-level annotation: we provide a self-supervised process reward model, which can directly learn the high-quality reasoning trajectory selection from the outcome reward. Equipped with the reflective generative form, MetaStone-S1 is naturally suitable for test-time scaling, and we provide three reasoning effort modes (low, medium, and high) based on the controllable thinking length. Experiments demonstrate that our MetaStone-S1 achieves comparable performance to OpenAI o3-mini's series with only 32B parameter size. To support the research community, we have open-sourced MetaStone-S1 at https://github.com/MetaStone-AI/MetaStone-S1.
♻ ☆ GuidedBench: Measuring and Mitigating the Evaluation Discrepancies of In-the-wild LLM Jailbreak Methods
Despite the growing interest in jailbreak methods as an effective red-teaming tool for building safe and responsible large language models (LLMs), flawed evaluation system designs have led to significant discrepancies in their effectiveness assessments. We conduct a systematic measurement study based on 37 jailbreak studies since 2022, focusing on both the methods and the evaluation systems they employ. We find that existing evaluation systems lack case-specific criteria, resulting in misleading conclusions about their effectiveness and safety implications. This paper advocates a shift to a more nuanced, case-by-case evaluation paradigm. We introduce GuidedBench, a novel benchmark comprising a curated harmful question dataset, detailed case-by-case evaluation guidelines and an evaluation system integrated with these guidelines -- GuidedEval. Experiments demonstrate that GuidedBench offers more accurate measurements of jailbreak performance, enabling meaningful comparisons across methods and uncovering new insights overlooked in previous evaluations. GuidedEval reduces inter-evaluator variance by at least 76.03\%. Furthermore, we observe that incorporating guidelines can enhance the effectiveness of jailbreak methods themselves, offering new insights into both attack strategies and evaluation paradigms.
comment: Homepage: https://sproutnan.github.io/AI-Safety_Benchmark/
Computer Vision and Pattern Recognition 6
☆ Entity Re-identification in Visual Storytelling via Contrastive Reinforcement Learning
Visual storytelling systems, particularly large vision-language models, struggle to maintain character and object identity across frames, often failing to recognize when entities in different images represent the same individuals or objects, leading to inconsistent references and referential hallucinations. This occurs because models lack explicit training on when to establish entity connections across frames. We propose a contrastive reinforcement learning approach that trains models to discriminate between coherent image sequences and stories from unrelated images. We extend the Story Reasoning dataset with synthetic negative examples to teach appropriate entity connection behavior. We employ Direct Preference Optimization with a dual-component reward function that promotes grounding and re-identification of entities in real stories while penalizing incorrect entity connections in synthetic contexts. Using this contrastive framework, we fine-tune Qwen Storyteller (based on Qwen2.5-VL 7B). Evaluation shows improvements in grounding mAP from 0.27 to 0.31 (+14.8%), F1 from 0.35 to 0.41 (+17.1%). Pronoun grounding accuracy improved across all pronoun types except ``its'', and cross-frame character and object persistence increased across all frame counts, with entities appearing in 5 or more frames advancing from 29.3% to 33.3% (+13.7%). Well-structured stories, containing the chain-of-thought and grounded story, increased from 79.1% to 97.5% (+23.3%).
comment: 7 pages
☆ Scalable and Realistic Virtual Try-on Application for Foundation Makeup with Kubelka-Munk Theory CVPR 2025
Augmented reality is revolutionizing beauty industry with virtual try-on (VTO) applications, which empowers users to try a wide variety of products using their phones without the hassle of physically putting on real products. A critical technical challenge in foundation VTO applications is the accurate synthesis of foundation-skin tone color blending while maintaining the scalability of the method across diverse product ranges. In this work, we propose a novel method to approximate well-established Kubelka-Munk (KM) theory for faster image synthesis while preserving foundation-skin tone color blending realism. Additionally, we build a scalable end-to-end framework for realistic foundation makeup VTO solely depending on the product information available on e-commerce sites. We validate our method using real-world makeup images, demonstrating that our framework outperforms other techniques.
comment: Presented at the workshop Three questions about virtual try-on at CVPR 2025
☆ mmFlux: Crowd Flow Analytics with Commodity mmWave MIMO Radar
In this paper, we present a novel framework for extracting underlying crowd motion patterns and inferring crowd semantics using mmWave radar. First, our proposed signal processing pipeline combines optical flow estimation concepts from vision with novel statistical and morphological noise filtering to generate high-fidelity mmWave flow fields - compact 2D vector representations of crowd motion. We then introduce a novel approach that transforms these fields into directed geometric graphs, where edges capture dominant flow currents, vertices mark crowd splitting or merging, and flow distribution is quantified across edges. Finally, we show that by analyzing the local Jacobian and computing the corresponding curl and divergence, we can extract key crowd semantics for both structured and diffused crowds. We conduct 21 experiments on crowds of up to (and including) 20 people across 3 areas, using commodity mmWave radar. Our framework achieves high-fidelity graph reconstruction of the underlying flow structure, even for complex crowd patterns, demonstrating strong spatial alignment and precise quantitative characterization of flow split ratios. Finally, our curl and divergence analysis accurately infers key crowd semantics, e.g., abrupt turns, boundaries where flow directions shift, dispersions, and gatherings. Overall, these findings validate our framework, underscoring its potential for various crowd analytics applications.
☆ ADIEE: Automatic Dataset Creation and Scorer for Instruction-Guided Image Editing Evaluation ICCV
Recent advances in instruction-guided image editing underscore the need for effective automated evaluation. While Vision-Language Models (VLMs) have been explored as judges, open-source models struggle with alignment, and proprietary models lack transparency and cost efficiency. Additionally, no public training datasets exist to fine-tune open-source VLMs, only small benchmarks with diverse evaluation schemes. To address this, we introduce ADIEE, an automated dataset creation approach which is then used to train a scoring model for instruction-guided image editing evaluation. We generate a large-scale dataset with over 100K samples and use it to fine-tune a LLaVA-NeXT-8B model modified to decode a numeric score from a custom token. The resulting scorer outperforms all open-source VLMs and Gemini-Pro 1.5 across all benchmarks, achieving a 0.0696 (+17.24%) gain in score correlation with human ratings on AURORA-Bench, and improving pair-wise comparison accuracy by 4.03% (+7.21%) on GenAI-Bench and 4.75% (+9.35%) on AURORA-Bench, respectively, compared to the state-of-the-art. The scorer can act as a reward model, enabling automated best edit selection and model fine-tuning. Notably, the proposed scorer can boost MagicBrush model's average evaluation score on ImagenHub from 5.90 to 6.43 (+8.98%).
comment: International Conference on Computer Vision (ICCV) 2025
♻ ☆ Adversarial Defenses via Vector Quantization
Adversarial attacks pose significant challenges to the robustness of modern deep neural networks in computer vision, and defending these networks against adversarial attacks has attracted intense research efforts. Among various defense strategies, preprocessing-based defenses are practically appealing since there is no need to train the network under protection. However, such approaches typically do not achieve comparable robustness as other methods such as adversarial training. In this paper, we propose a novel framework for preprocessing-based defenses, where a vector quantizer is used as a preprocessor. This framework, inspired by and extended from Randomized Discretization (RandDisc), is theoretically principled by rate-distortion theory: indeed, RandDisc may be viewed as a scalar quantizer, and rate-distortion theory suggests that such quantization schemes are inferior to vector quantization. In our framework, the preprocessing vector quantizer treats the input image as a collection of patches and finds a set of representative patches based on the patch distributions; each original patch is then modified according to the representative patches close to it. We present two lightweight defenses in this framework, referred to as patched RandDisc (pRD) and sliding-window RandDisc (swRD), where the patches are disjoint in the former and overlapping in the latter. We show that vector-quantization-based defenses have certifiable robust accuracy and that pRD and swRD demonstrate state-of-the-art performances, surpassing RandDisc by a large margin. Notably, the proposed defenses possess the obfuscated gradients property. Our experiments however show that pRD and swRD remain effective under the STE and EOT attacks, which are designed specifically for defenses with gradient obfuscation. ...
comment: This is the author-accepted version of our paper published in Neurocomputing. The final published version is available at: https://doi.org/10.1016/j.neucom.2025.130703
♻ ☆ Taming the Tri-Space Tension: ARC-Guided Hallucination Modeling and Control for Text-to-Image Generation
Despite remarkable progress in image quality and prompt fidelity, text-to-image (T2I) diffusion models continue to exhibit persistent "hallucinations", where generated content subtly or significantly diverges from the intended prompt semantics. While often regarded as unpredictable artifacts, we argue that these failures reflect deeper, structured misalignments within the generative process. In this work, we propose a cognitively inspired perspective that reinterprets hallucinations as trajectory drift within a latent alignment space. Empirical observations reveal that generation unfolds within a multiaxial cognitive tension field, where the model must continuously negotiate competing demands across three key critical axes: semantic coherence, structural alignment, and knowledge grounding. We then formalize this three-axis space as the \textbf{Hallucination Tri-Space} and introduce the Alignment Risk Code (ARC): a dynamic vector representation that quantifies real-time alignment tension during generation. The magnitude of ARC captures overall misalignment, its direction identifies the dominant failure axis, and its imbalance reflects tension asymmetry. Based on this formulation, we develop the TensionModulator (TM-ARC): a lightweight controller that operates entirely in latent space. TM-ARC monitors ARC signals and applies targeted, axis-specific interventions during the sampling process. Extensive experiments on standard T2I benchmarks demonstrate that our approach significantly reduces hallucination without compromising image quality or diversity. This framework offers a unified and interpretable approach for understanding and mitigating generative failures in diffusion-based T2I systems.
comment: We withdraw this paper due to significant visualization errors in Figure 3 and 5 that affect the correctness of our core modeling claims and may cause misinterpretation. These figures misrepresent ARC dynamics and trajectory control
Information Retrieval 17
☆ A Language-Driven Framework for Improving Personalized Recommendations: Merging LLMs with Traditional Algorithms
Traditional recommendation algorithms are not designed to provide personalized recommendations based on user preferences provided through text, e.g., "I enjoy light-hearted comedies with a lot of humor". Large Language Models (LLMs) have emerged as one of the most promising tools for natural language processing in recent years. This research proposes a novel framework that mimics how a close friend would recommend items based on their knowledge of an individual's tastes. We leverage LLMs to enhance movie recommendation systems by refining traditional algorithm outputs and integrating them with language-based user preference inputs. We employ Singular Value Decomposition (SVD) or SVD++ algorithms to generate initial movie recommendations, implemented using the Surprise Python library and trained on the MovieLens-Latest-Small dataset. We compare the performance of the base algorithms with our LLM-enhanced versions using leave-one-out validation hit rates and cumulative hit rates. Additionally, to compare the performance of our framework against the current state-of-the-art recommendation systems, we use rating and ranking metrics with an item-based stratified 0.75 train, 0.25 test split. Our framework can generate preference profiles automatically based on users' favorite movies or allow manual preference specification for more personalized results. Using an automated approach, our framework overwhelmingly surpassed SVD and SVD++ on every evaluation metric used (e.g., improvements of up to ~6x in cumulative hit rate, ~3.7x in NDCG, etc.), albeit at the cost of a slight increase in computational overhead.
☆ Boosting Parameter Efficiency in LLM-Based Recommendation through Sophisticated Pruning
LLM-based recommender systems have made significant progress; however, the deployment cost associated with the large parameter volume of LLMs still hinders their real-world applications. This work explores parameter pruning to improve parameter efficiency while maintaining recommendation quality, thereby enabling easier deployment. Unlike existing approaches that focus primarily on inter-layer redundancy, we uncover intra-layer redundancy within components such as self-attention and MLP modules. Building on this analysis, we propose a more fine-grained pruning approach that integrates both intra-layer and layer-wise pruning. Specifically, we introduce a three-stage pruning strategy that progressively prunes parameters at different levels and parts of the model, moving from intra-layer to layer-wise pruning, or from width to depth. Each stage also includes a performance restoration step using distillation techniques, helping to strike a balance between performance and parameter efficiency. Empirical results demonstrate the effectiveness of our approach: across three datasets, our models achieve an average of 88% of the original model's performance while pruning more than 95% of the non-embedding parameters. This underscores the potential of our method to significantly reduce resource requirements without greatly compromising recommendation quality. Our code will be available at: https://github.com/zheng-sl/PruneRec
☆ SCoRE: Streamlined Corpus-based Relation Extraction using Multi-Label Contrastive Learning and Bayesian kNN
The growing demand for efficient knowledge graph (KG) enrichment leveraging external corpora has intensified interest in relation extraction (RE), particularly under low-supervision settings. To address the need for adaptable and noise-resilient RE solutions that integrate seamlessly with pre-trained large language models (PLMs), we introduce SCoRE, a modular and cost-effective sentence-level RE system. SCoRE enables easy PLM switching, requires no finetuning, and adapts smoothly to diverse corpora and KGs. By combining supervised contrastive learning with a Bayesian k-Nearest Neighbors (kNN) classifier for multi-label classification, it delivers robust performance despite the noisy annotations of distantly supervised corpora. To improve RE evaluation, we propose two novel metrics: Correlation Structure Distance (CSD), measuring the alignment between learned relational patterns and KG structures, and Precision at R (P@R), assessing utility as a recommender system. We also release Wiki20d, a benchmark dataset replicating real-world RE conditions where only KG-derived annotations are available. Experiments on five benchmarks show that SCoRE matches or surpasses state-of-the-art methods while significantly reducing energy consumption. Further analyses reveal that increasing model complexity, as seen in prior work, degrades performance, highlighting the advantages of SCoRE's minimal design. Combining efficiency, modularity, and scalability, SCoRE stands as an optimal choice for real-world RE applications.
☆ CDC: Causal Domain Clustering for Multi-Domain Recommendation SIGIR 2025
Multi-domain recommendation leverages domain-general knowledge to improve recommendations across several domains. However, as platforms expand to dozens or hundreds of scenarios, training all domains in a unified model leads to performance degradation due to significant inter-domain differences. Existing domain grouping methods, based on business logic or data similarities, often fail to capture the true transfer relationships required for optimal grouping. To effectively cluster domains, we propose Causal Domain Clustering (CDC). CDC models domain transfer patterns within a large number of domains using two distinct effects: the Isolated Domain Affinity Matrix for modeling non-interactive domain transfers, and the Hybrid Domain Affinity Matrix for considering dynamic domain synergy or interference under joint training. To integrate these two transfer effects, we introduce causal discovery to calculate a cohesion-based coefficient that adaptively balances their contributions. A Co-Optimized Dynamic Clustering algorithm iteratively optimizes target domain clustering and source domain selection for training. CDC significantly enhances performance across over 50 domains on public datasets and in industrial settings, achieving a 4.9% increase in online eCPM. Code is available at https://github.com/Chrissie-Law/Causal-Domain-Clustering-for-Multi-Domain-Recommendation
comment: Accepted at SIGIR 2025
☆ Temporal Information Retrieval via Time-Specifier Model Merging
The rapid expansion of digital information and knowledge across structured and unstructured sources has heightened the importance of Information Retrieval (IR). While dense retrieval methods have substantially improved semantic matching for general queries, they consistently underperform on queries with explicit temporal constraints--often those containing numerical expressions and time specifiers such as ``in 2015.'' Existing approaches to Temporal Information Retrieval (TIR) improve temporal reasoning but often suffer from catastrophic forgetting, leading to reduced performance on non-temporal queries. To address this, we propose Time-Specifier Model Merging (TSM), a novel method that enhances temporal retrieval while preserving accuracy on non-temporal queries. TSM trains specialized retrievers for individual time specifiers and merges them in to a unified model, enabling precise handling of temporal constraints without compromising non-temporal retrieval. Extensive experiments on both temporal and non-temporal datasets demonstrate that TSM significantly improves performance on temporally constrained queries while maintaining strong results on non-temporal queries, consistently outperforming other baseline methods. Our code is available at https://github.com/seungyoonee/TSM .
☆ CLI-RAG: A Retrieval-Augmented Framework for Clinically Structured and Context Aware Text Generation with LLMs
Large language models (LLMs), including zero-shot and few-shot paradigms, have shown promising capabilities in clinical text generation. However, real-world applications face two key challenges: (1) patient data is highly unstructured, heterogeneous, and scattered across multiple note types and (2) clinical notes are often long and semantically dense, making naive prompting infeasible due to context length constraints and the risk of omitting clinically relevant information. We introduce CLI-RAG (Clinically Informed Retrieval-Augmented Generation), a domain-specific framework for structured and clinically grounded text generation using LLMs. It incorporates a novel hierarchical chunking strategy that respects clinical document structure and introduces a task-specific dual-stage retrieval mechanism. The global stage identifies relevant note types using evidence-based queries, while the local stage extracts high-value content within those notes creating relevance at both document and section levels. We apply the system to generate structured progress notes for individual hospital visits using 15 clinical note types from the MIMIC-III dataset. Experiments show that it preserves temporal and semantic alignment across visits, achieving an average alignment score of 87.7%, surpassing the 80.7% baseline from real clinician-authored notes. The generated outputs also demonstrate high consistency across LLMs, reinforcing deterministic behavior essential for reproducibility, reliability, and clinical trust.
comment: 12 pages, 4 figures
☆ MS-DPPs: Multi-Source Determinantal Point Processes for Contextual Diversity Refinement of Composite Attributes in Text to Image Retrieval IJCAI 2025
Result diversification (RD) is a crucial technique in Text-to-Image Retrieval for enhancing the efficiency of a practical application. Conventional methods focus solely on increasing the diversity metric of image appearances. However, the diversity metric and its desired value vary depending on the application, which limits the applications of RD. This paper proposes a novel task called CDR-CA (Contextual Diversity Refinement of Composite Attributes). CDR-CA aims to refine the diversities of multiple attributes, according to the application's context. To address this task, we propose Multi-Source DPPs, a simple yet strong baseline that extends the Determinantal Point Process (DPP) to multi-sources. We model MS-DPP as a single DPP model with a unified similarity matrix based on a manifold representation. We also introduce Tangent Normalization to reflect contexts. Extensive experiments demonstrate the effectiveness of the proposed method. Our code is publicly available at https://github.com/NEC-N-SOGI/msdpp.
comment: IJCAI 2025. Code: https://github.com/NEC-N-SOGI/msdpp
☆ Impacts of Mainstream-Driven Algorithms on Recommendations for Children Across Domains: A Reproducibility Study RecSys 2025
Children are often exposed to items curated by recommendation algorithms. Yet, research seldom considers children as a user group, and when it does, it is anchored on datasets where children are underrepresented, risking overlooking their interests, favoring those of the majority, i.e., mainstream users. Recently, Ungruh et al. demonstrated that children's consumption patterns and preferences differ from those of mainstream users, resulting in inconsistent recommendation algorithm performance and behavior for this user group. These findings, however, are based on two datasets with a limited child user sample. We reproduce and replicate this study on a wider range of datasets in the movie, music, and book domains, uncovering interaction patterns and aspects of child-recommender interactions consistent across domains, as well as those specific to some user samples in the data. We also extend insights from the original study with popularity bias metrics, given the interpretation of results from the original study. With this reproduction and extension, we uncover consumption patterns and differences between age groups stemming from intrinsic differences between children and others, and those unique to specific datasets or domains.
comment: Preprint of accepted RecSys 2025 contribution
☆ DS@GT at CheckThat! 2025: Exploring Retrieval and Reranking Pipelines for Scientific Claim Source Retrieval on Social Media Discourse
Social media users often make scientific claims without citing where these claims come from, generating a need to verify these claims. This paper details work done by the DS@GT team for CLEF 2025 CheckThat! Lab Task 4b Scientific Claim Source Retrieval which seeks to find relevant scientific papers based on implicit references in tweets. Our team explored 6 different data augmentation techniques, 7 different retrieval and reranking pipelines, and finetuned a bi-encoder. Achieving an MRR@5 of 0.58, our team ranked 16th out of 30 teams for the CLEF 2025 CheckThat! Lab Task 4b, and improvement of 0.15 over the BM25 baseline of 0.43. Our code is available on Github at https://github.com/dsgt-arc/checkthat-2025-swd/tree/main/subtask-4b.
☆ SPEAR: Subset-sampled Performance Evaluation via Automated Ground Truth Generation for RAG
Retrieval-Augmented Generation (RAG) is a core approach for enhancing Large Language Models (LLMs), where the effectiveness of the retriever largely determines the overall response quality of RAG systems. Retrievers encompass a multitude of hyperparameters that significantly impact performance outcomes and demonstrate sensitivity to specific applications. Nevertheless, hyperparameter optimization entails prohibitively high computational expenses. Existing evaluation methods suffer from either prohibitive costs or disconnection from domain-specific scenarios. This paper proposes SEARA (Subset sampling Evaluation for Automatic Retriever Assessment), which addresses evaluation data challenges through subset sampling techniques and achieves robust automated retriever evaluation by minimal retrieval facts extraction and comprehensive retrieval metrics. Based on real user queries, this method enables fully automated retriever evaluation at low cost, thereby obtaining optimal retriever for specific business scenarios. We validate our method across classic RAG applications in rednote, including knowledge-based Q&A system and retrieval-based travel assistant, successfully obtaining scenario-specific optimal retrievers.
☆ GR-LLMs: Recent Advances in Generative Recommendation Based on Large Language Models
In the past year, Generative Recommendations (GRs) have undergone substantial advancements, especially in leveraging the powerful sequence modeling and reasoning capabilities of Large Language Models (LLMs) to enhance overall recommendation performance. LLM-based GRs are forming a new paradigm that is distinctly different from discriminative recommendations, showing strong potential to replace traditional recommendation systems heavily dependent on complex hand-crafted features. In this paper, we provide a comprehensive survey aimed at facilitating further research of LLM-based GRs. Initially, we outline the general preliminaries and application cases of LLM-based GRs. Subsequently, we introduce the main considerations when LLM-based GRs are applied in real industrial scenarios. Finally, we explore promising directions for LLM-based GRs. We hope that this survey contributes to the ongoing advancement of the GR domain.
comment: 8 pages, 3 figures
☆ UniConv: Unifying Retrieval and Response Generation for Large Language Models in Conversations ACL 2025
The rapid advancement of conversational search systems revolutionizes how information is accessed by enabling the multi-turn interaction between the user and the system. Existing conversational search systems are usually built with two different models. This separation restricts the system from leveraging the intrinsic knowledge of the models simultaneously, which cannot ensure the effectiveness of retrieval benefiting the generation. The existing studies for developing unified models cannot fully address the aspects of understanding conversational context, managing retrieval independently, and generating responses. In this paper, we explore how to unify dense retrieval and response generation for large language models in conversation. We conduct joint fine-tuning with different objectives and design two mechanisms to reduce the inconsistency risks while mitigating data discrepancy. The evaluations on five conversational search datasets demonstrate that our unified model can mutually improve both tasks and outperform the existing baselines.
comment: Accepted by ACL 2025 (main)
♻ ☆ Rankers, Judges, and Assistants: Towards Understanding the Interplay of LLMs in Information Retrieval Evaluation SIGIR
Large language models (LLMs) are increasingly integral to information retrieval (IR), powering ranking, evaluation, and AI-assisted content creation. This widespread adoption necessitates a critical examination of potential biases arising from the interplay between these LLM-based components. This paper synthesizes existing research and presents novel experiment designs that explore how LLM-based rankers and assistants influence LLM-based judges. We provide the first empirical evidence of LLM judges exhibiting significant bias towards LLM-based rankers. Furthermore, we observe limitations in LLM judges' ability to discern subtle system performance differences. Contrary to some previous findings, our preliminary study does not find evidence of bias against AI-generated content. These results highlight the need for a more holistic view of the LLM-driven information ecosystem. To this end, we offer initial guidelines and a research agenda to ensure the reliable use of LLMs in IR evaluation.
comment: Proceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval (SIGIR '25)
♻ ☆ IntOPE: Off-Policy Evaluation in the Presence of Interference
Off-Policy Evaluation (OPE) is employed to assess the potential impact of a hypothetical policy using logged contextual bandit feedback, which is crucial in areas such as personalized medicine and recommender systems, where online interactions are associated with significant risks and costs. Traditionally, OPE methods rely on the Stable Unit Treatment Value Assumption (SUTVA), which assumes that the reward for any given individual is unaffected by the actions of others. However, this assumption often fails in real-world scenarios due to the presence of interference, where an individual's reward is affected not just by their own actions but also by the actions of their peers. This realization reveals significant limitations of existing OPE methods in real-world applications. To address this limitation, we propose IntIPW, an IPW-style estimator that extends the Inverse Probability Weighting (IPW) framework by integrating marginalized importance weights to account for both individual actions and the influence of adjacent entities. Extensive experiments are conducted on both synthetic and real-world data to demonstrate the effectiveness of the proposed IntIPW method.
♻ ☆ Multi-task Offline Reinforcement Learning for Online Advertising in Recommender Systems KDD 2025
Online advertising in recommendation platforms has gained significant attention, with a predominant focus on channel recommendation and budget allocation strategies. However, current offline reinforcement learning (RL) methods face substantial challenges when applied to sparse advertising scenarios, primarily due to severe overestimation, distributional shifts, and overlooking budget constraints. To address these issues, we propose MTORL, a novel multi-task offline RL model that targets two key objectives. First, we establish a Markov Decision Process (MDP) framework specific to the nuances of advertising. Then, we develop a causal state encoder to capture dynamic user interests and temporal dependencies, facilitating offline RL through conditional sequence modeling. Causal attention mechanisms are introduced to enhance user sequence representations by identifying correlations among causal states. We employ multi-task learning to decode actions and rewards, simultaneously addressing channel recommendation and budget allocation. Notably, our framework includes an automated system for integrating these tasks into online advertising. Extensive experiments on offline and online environments demonstrate MTORL's superiority over state-of-the-art methods.
comment: KDD 2025
♻ ☆ FinSphere, a Real-Time Stock Analysis Agent Powered by Instruction-Tuned LLMs and Domain Tools
Current financial large language models (FinLLMs) struggle with two critical limitations: the absence of objective evaluation metrics to assess the quality of stock analysis reports and a lack of depth in stock analysis, which impedes their ability to generate professional-grade insights. To address these challenges, this paper introduces FinSphere, a stock analysis agent, along with three major contributions: (1) AnalyScore, a systematic evaluation framework for assessing stock analysis quality, (2) Stocksis, a dataset curated by industry experts to enhance LLMs' stock analysis capabilities, and (3) FinSphere, an AI agent that can generate high-quality stock analysis reports in response to user queries. Experiments demonstrate that FinSphere achieves superior performance compared to both general and domain-specific LLMs, as well as existing agent-based systems, even when they are enhanced with real-time data access and few-shot guidance. The integrated framework, which combines real-time data feeds, quantitative tools, and an instruction-tuned LLM, yields substantial improvements in both analytical quality and practical applicability for real-world stock analysis.
♻ ☆ Hespi: A pipeline for automatically detecting information from hebarium specimen sheets
Specimen-associated biodiversity data are crucial for biological, environmental, and conservation sciences. A rate shift is needed to extract data from specimen images efficiently, moving beyond human-mediated transcription. We developed `Hespi' (HErbarium Specimen sheet PIpeline) using advanced computer vision techniques to extract pre-catalogue data from primary specimen labels on herbarium specimens. Hespi integrates two object detection models: one for detecting the components of the sheet and another for fields on the primary primary specimen label. It classifies labels as printed, typed, handwritten, or mixed and uses Optical Character Recognition (OCR) and Handwritten Text Recognition (HTR) for extraction. The text is then corrected against authoritative taxon databases and refined using a multimodal Large Language Model (LLM). Hespi accurately detects and extracts text from specimen sheets across international herbaria, and its modular design allows users to train and integrate custom models.
comment: 15 pages, 7 figures
Multimedia 8
☆ Audio-Visual Speech Separation via Bottleneck Iterative Network
Integration of information from non-auditory cues can significantly improve the performance of speech-separation models. Often such models use deep modality-specific networks to obtain unimodal features, and risk being too costly or lightweight but lacking capacity. In this work, we present an iterative representation refinement approach called Bottleneck Iterative Network (BIN), a technique that repeatedly progresses through a lightweight fusion block, while bottlenecking fusion representations by fusion tokens. This helps improve the capacity of the model, while avoiding major increase in model size and balancing between the model performance and training cost. We test BIN on challenging noisy audio-visual speech separation tasks, and show that our approach consistently outperforms state-of-the-art benchmark models with respect to SI-SDRi on NTCD-TIMIT and LRS3+WHAM! datasets, while simultaneously achieving a reduction of more than 50% in training and GPU inference time across nearly all settings.
comment: Accepted to the 42nd International Conference on Machine Learning Workshop on Machine Learning for Audio
☆ Semi-fragile watermarking of remote sensing images using DWT, vector quantization and automatic tiling
A semi-fragile watermarking scheme for multiple band images is presented in this article. We propose to embed a mark into remote sensing images applying a tree-structured vector quantization approach to the pixel signatures instead of processing each band separately. The signature of the multispectral or hyperspectral image is used to embed the mark in it order to detect any significant modification of the original image. The image is segmented into three-dimensional blocks, and a tree-structured vector quantizer is built for each block. These trees are manipulated using an iterative algorithm until the resulting block satisfies a required criterion, which establishes the embedded mark. The method is shown to be able to preserve the mark under lossy compression (above a given threshold) but, at the same time, it detects possibly forged blocks and their position in the whole image.
☆ MST-Distill: Mixture of Specialized Teachers for Cross-Modal Knowledge Distillation ACM MM 2025
Knowledge distillation as an efficient knowledge transfer technique, has achieved remarkable success in unimodal scenarios. However, in cross-modal settings, conventional distillation methods encounter significant challenges due to data and statistical heterogeneities, failing to leverage the complementary prior knowledge embedded in cross-modal teacher models. This paper empirically reveals two critical issues in existing approaches: distillation path selection and knowledge drift. To address these limitations, we propose MST-Distill, a novel cross-modal knowledge distillation framework featuring a mixture of specialized teachers. Our approach employs a diverse ensemble of teacher models across both cross-modal and multimodal configurations, integrated with an instance-level routing network that facilitates adaptive and dynamic distillation. This architecture effectively transcends the constraints of traditional methods that rely on monotonous and static teacher models. Additionally, we introduce a plug-in masking module, independently trained to suppress modality-specific discrepancies and reconstruct teacher representations, thereby mitigating knowledge drift and enhancing transfer effectiveness. Extensive experiments across five diverse multimodal datasets, spanning visual, audio, and text, demonstrate that our method significantly outperforms existing state-of-the-art knowledge distillation methods in cross-modal distillation tasks. The source code is available at https://github.com/Gray-OREO/MST-Distill.
comment: Accepted to ACM MM 2025 (The 33rd ACM International Conference on Multimedia)
☆ Dual-Granularity Cross-Modal Identity Association for Weakly-Supervised Text-to-Person Image Matching
Weakly supervised text-to-person image matching, as a crucial approach to reducing models' reliance on large-scale manually labeled samples, holds significant research value. However, existing methods struggle to predict complex one-to-many identity relationships, severely limiting performance improvements. To address this challenge, we propose a local-and-global dual-granularity identity association mechanism. Specifically, at the local level, we explicitly establish cross-modal identity relationships within a batch, reinforcing identity constraints across different modalities and enabling the model to better capture subtle differences and correlations. At the global level, we construct a dynamic cross-modal identity association network with the visual modality as the anchor and introduce a confidence-based dynamic adjustment mechanism, effectively enhancing the model's ability to identify weakly associated samples while improving overall sensitivity. Additionally, we propose an information-asymmetric sample pair construction method combined with consistency learning to tackle hard sample mining and enhance model robustness. Experimental results demonstrate that the proposed method substantially boosts cross-modal matching accuracy, providing an efficient and practical solution for text-to-person image matching.
☆ Residual Prior-driven Frequency-aware Network for Image Fusion
Image fusion aims to integrate complementary information across modalities to generate high-quality fused images, thereby enhancing the performance of high-level vision tasks. While global spatial modeling mechanisms show promising results, constructing long-range feature dependencies in the spatial domain incurs substantial computational costs. Additionally, the absence of ground-truth exacerbates the difficulty of capturing complementary features effectively. To tackle these challenges, we propose a Residual Prior-driven Frequency-aware Network, termed as RPFNet. Specifically, RPFNet employs a dual-branch feature extraction framework: the Residual Prior Module (RPM) extracts modality-specific difference information from residual maps, thereby providing complementary priors for fusion; the Frequency Domain Fusion Module (FDFM) achieves efficient global feature modeling and integration through frequency-domain convolution. Additionally, the Cross Promotion Module (CPM) enhances the synergistic perception of local details and global structures through bidirectional feature interaction. During training, we incorporate an auxiliary decoder and saliency structure loss to strengthen the model's sensitivity to modality-specific differences. Furthermore, a combination of adaptive weight-based frequency contrastive loss and SSIM loss effectively constrains the solution space, facilitating the joint capture of local details and global features while ensuring the retention of complementary information. Extensive experiments validate the fusion performance of RPFNet, which effectively integrates discriminative features, enhances texture details and salient objects, and can effectively facilitate the deployment of the high-level vision task.
☆ QoE Optimization for Semantic Self-Correcting Video Transmission in Multi-UAV Networks
Real-time unmanned aerial vehicle (UAV) video streaming is essential for time-sensitive applications, including remote surveillance, emergency response, and environmental monitoring. However, it faces challenges such as limited bandwidth, latency fluctuations, and high packet loss. To address these issues, we propose a novel semantic self-correcting video transmission framework with ultra-fine bitrate granularity (SSCV-G). In SSCV-G, video frames are encoded into a compact semantic codebook space, and the transmitter adaptively sends a subset of semantic indices based on bandwidth availability, enabling fine-grained bitrate control for improved bandwidth efficiency. At the receiver, a spatio-temporal vision transformer (ST-ViT) performs multi-frame joint decoding to reconstruct dropped semantic indices by modeling intra- and inter-frame dependencies. To further improve performance under dynamic network conditions, we integrate a multi-user proximal policy optimization (MUPPO) reinforcement learning scheme that jointly optimizes communication resource allocation and semantic bitrate selection to maximize user Quality of Experience (QoE). Extensive experiments demonstrate that the proposed SSCV-G significantly outperforms state-of-the-art video codecs in coding efficiency, bandwidth adaptability, and packet loss robustness. Moreover, the proposed MUPPO-based QoE optimization consistently surpasses existing benchmarks.
comment: 13 pages
♻ ☆ Reply with Sticker: New Dataset and Model for Sticker Retrieval
Using stickers in online chatting is very prevalent on social media platforms, where the stickers used in the conversation can express someone's intention/emotion/attitude in a vivid, tactful, and intuitive way. Existing sticker retrieval research typically retrieves stickers based on context and the current utterance delivered by the user. That is, the stickers serve as a supplement to the current utterance. However, in the real-world scenario, using stickers to express what we want to say rather than as a supplement to our words only is also important. Therefore, in this paper, we create a new dataset for sticker retrieval in conversation, called \textbf{StickerInt}, where stickers are used to reply to previous conversations or supplement our words. Based on the created dataset, we present a simple yet effective framework for sticker retrieval in conversation based on the learning of intention and the cross-modal relationships between conversation context and stickers, coined as \textbf{Int-RA}. Specifically, we first devise a knowledge-enhanced intention predictor to introduce the intention information into the conversation representations. Subsequently, a relation-aware sticker selector is devised to retrieve the response sticker via cross-modal relationships. Extensive experiments on two datasets show that the proposed model achieves state-of-the-art performance and generalization capability in sticker retrieval. The dataset and source code of this work are released at https://github.com/HITSZ-HLT/Int-RA.
♻ ☆ UniF$^2$ace: Fine-grained Face Understanding and Generation with Unified Multimodal Models
Unified multimodal models (UMMs) have emerged as a powerful paradigm in foundational computer vision research, demonstrating significant potential in both image understanding and generation. However, existing research in the face domain primarily focuses on $\textbf{coarse}$ facial attribute understanding, with limited capacity to handle $\textbf{fine-grained}$ facial attributes and without addressing generation capabilities. To overcome these limitations, we propose UniF$^2$ace, the first UMM tailored specifically for fine-grained face understanding and generation. In general, we train UniF$^2$ace on a self-constructed, specialized dataset utilizing two mutually beneficial diffusion techniques and a two-level mixture-of-experts architecture. Specifically, we first build a large-scale facial dataset, UniF$^2$ace-130K, which contains 130K image-text pairs with one million question-answering pairs that span a wide range of facial attributes. Second, we establish a theoretical connection between discrete diffusion score matching and masked generative models, optimizing both evidence lower bounds simultaneously, which significantly improves the model's ability to synthesize facial details. Finally, we introduce both token-level and sequence-level mixture-of-experts, enabling efficient fine-grained representation learning for both understanding and generation tasks. Extensive experiments on UniF$^2$ace-130K demonstrate that UniF$^2$ace outperforms existing UMMs and generative models, achieving superior performance across both understanding and generation tasks.
Information Retrieval 22
☆ Unconditional Diffusion for Generative Sequential Recommendation
Diffusion models, known for their generative ability to simulate data creation through noise-adding and denoising processes, have emerged as a promising approach for building generative recommenders. To incorporate user history for personalization, existing methods typically adopt a conditional diffusion framework, where the reverse denoising process of reconstructing items from noise is modified to be conditioned on the user history. However, this design may fail to fully utilize historical information, as it gets distracted by the need to model the "item $\leftrightarrow$ noise" translation. This motivates us to reformulate the diffusion process for sequential recommendation in an unconditional manner, treating user history (instead of noise) as the endpoint of the forward diffusion process (i.e., the starting point of the reverse process), rather than as a conditional input. This formulation allows for exclusive focus on modeling the "item $\leftrightarrow$ history" translation. To this end, we introduce Brownian Bridge Diffusion Recommendation (BBDRec). By leveraging a Brownian bridge process, BBDRec enforces a structured noise addition and denoising mechanism, ensuring that the trajectories are constrained towards a specific endpoint -- user history, rather than noise. Extensive experiments demonstrate BBDRec's effectiveness in enhancing sequential recommendation performance. The source code is available at https://github.com/baiyimeng/BBDRec.
☆ Tile-Based ViT Inference with Visual-Cluster Priors for Zero-Shot Multi-Species Plant Identification
We describe DS@GT's second-place solution to the PlantCLEF 2025 challenge on multi-species plant identification in vegetation quadrat images. Our pipeline combines (i) a fine-tuned Vision Transformer ViTD2PC24All for patch-level inference, (ii) a 4x4 tiling strategy that aligns patch size with the network's 518x518 receptive field, and (iii) domain-prior adaptation through PaCMAP + K-Means visual clustering and geolocation filtering. Tile predictions are aggregated by majority vote and re-weighted with cluster-specific Bayesian priors, yielding a macro-averaged F1 of 0.348 (private leaderboard) while requiring no additional training. All code, configuration files, and reproducibility scripts are publicly available at https://github.com/dsgt-arc/plantclef-2025.
☆ Nyay-Darpan: Enhancing Decision Making Through Summarization and Case Retrieval for Consumer Law in India
AI-based judicial assistance and case prediction have been extensively studied in criminal and civil domains, but remain largely unexplored in consumer law, especially in India. In this paper, we present Nyay-Darpan, a novel two-in-one framework that (i) summarizes consumer case files and (ii) retrieves similar case judgements to aid decision-making in consumer dispute resolution. Our methodology not only addresses the gap in consumer law AI tools but also introduces an innovative approach to evaluate the quality of the summary. The term 'Nyay-Darpan' translates into 'Mirror of Justice', symbolizing the ability of our tool to reflect the core of consumer disputes through precise summarization and intelligent case retrieval. Our system achieves over 75 percent accuracy in similar case prediction and approximately 70 percent accuracy across material summary evaluation metrics, demonstrating its practical effectiveness. We will publicly release the Nyay-Darpan framework and dataset to promote reproducibility and facilitate further research in this underexplored yet impactful domain.
☆ Contrastive and Transfer Learning for Effective Audio Fingerprinting through a Real-World Evaluation Protocol
Recent advances in song identification leverage deep neural networks to learn compact audio fingerprints directly from raw waveforms. While these methods perform well under controlled conditions, their accuracy drops significantly in real-world scenarios where the audio is captured via mobile devices in noisy environments. In this paper, we introduce a novel evaluation protocol designed to better reflect such real-world conditions. We generate three recordings of the same audio, each with increasing levels of noise, captured using a mobile device's microphone. Our results reveal a substantial performance drop for two state-of-the-art CNN-based models under this protocol, compared to previously reported benchmarks. Additionally, we highlight the critical role of the augmentation pipeline during training with contrastive loss. By introduction low pass and high pass filters in the augmentation pipeline we significantly increase the performance of both systems in our proposed evaluation. Furthermore, we develop a transformer-based model with a tailored projection module and demonstrate that transferring knowledge from a semantically relevant domain yields a more robust solution. The transformer architecture outperforms CNN-based models across all noise levels, and query durations. In low noise conditions it achieves 47.99% for 1-sec queries, and 97% for 10-sec queries in finding the correct song, surpassing by 14%, and by 18.5% the second-best performing model, respectively, Under heavy noise levels, we achieve a detection rate 56.5% for 15-second query duration. All experiments are conducted on public large-scale dataset of over 100K songs, with queries matched against a database of 56 million vectors.
comment: International Journal of Music Science, Technology and Art, 15 pages, 7 figures
☆ Hierarchical Interaction Summarization and Contrastive Prompting for Explainable Recommendations
Explainable recommendations, which use the information of user and item with interaction to generate a explanation for why the user would interact with the item, are crucial for improving user trust and decision transparency to the recommender system. Existing methods primarily rely on encoding features of users and items to embeddings, which often leads to information loss due to dimensionality reduction, sparse interactions, and so on. With the advancements of large language models (LLMs) in language comprehension, some methods use embeddings as LLM inputs for explanation generation. However, since embeddings lack inherent semantics, LLMs must adjust or extend their parameters to interpret them, a process that inevitably incurs information loss. To address this issue, we propose a novel approach combining profile generation via hierarchical interaction summarization (PGHIS), which leverages a pretrained LLM to hierarchically summarize user-item interactions, generating structured textual profiles as explicit representations of user and item characteristics. Additionally, we propose contrastive prompting for explanation generation (CPEG) which employs contrastive learning to guide another reasoning language models in producing high-quality ground truth recommendation explanations. Finally, we use the textual profiles of user and item as input and high-quality explanation as output to fine-tune a LLM for generating explanations. Experimental results on multiple datasets demonstrate that our approach outperforms existing state-of-the-art methods, achieving a great improvement on metrics about explainability (e.g., 5% on GPTScore) and text quality. Furthermore, our generated ground truth explanations achieve a significantly higher win rate compared to user-written reviews and those produced by other methods, demonstrating the effectiveness of CPEG in generating high-quality ground truths.
☆ Enhancing the Interpretability of Rule-based Explanations through Information Retrieval
The lack of transparency of data-driven Artificial Intelligence techniques limits their interpretability and acceptance into healthcare decision-making processes. We propose an attribution-based approach to improve the interpretability of Explainable AI-based predictions in the specific context of arm lymphedema's risk assessment after lymph nodal radiotherapy in breast cancer. The proposed method performs a statistical analysis of the attributes in the rule-based prediction model using standard metrics from Information Retrieval techniques. This analysis computes the relevance of each attribute to the prediction and provides users with interpretable information about the impact of risk factors. The results of a user study that compared the output generated by the proposed approach with the raw output of the Explainable AI model suggested higher levels of interpretability and usefulness in the context of predicting lymphedema risk.
☆ Semantic Certainty Assessment in Vector Retrieval Systems: A Novel Framework for Embedding Quality Evaluation
Vector retrieval systems exhibit significant performance variance across queries due to heterogeneous embedding quality. We propose a lightweight framework for predicting retrieval performance at the query level by combining quantization robustness and neighborhood density metrics. Our approach is motivated by the observation that high-quality embeddings occupy geometrically stable regions in the embedding space and exhibit consistent neighborhood structures. We evaluate our method on 4 standard retrieval datasets, showing consistent improvements of 9.4$\pm$1.2\% in Recall@10 over competitive baselines. The framework requires minimal computational overhead (less than 5\% of retrieval time) and enables adaptive retrieval strategies. Our analysis reveals systematic patterns in embedding quality across different query types, providing insights for targeted training data augmentation.
comment: 7 pages
☆ RecRankerEval: A Flexible and Extensible Framework for Top-k LLM-based Recommendation
A recent Large language model (LLM)-based recommendation model, called RecRanker, has demonstrated a superior performance in the top-k recommendation task compared to other models. In particular, RecRanker samples users via clustering, generates an initial ranking list using an initial recommendation model, and fine-tunes an LLM through hybrid instruction tuning to infer user preferences. However, the contribution of each core component remains underexplored. In this work, we inspect the reproducibility of RecRanker, and study the impact and role of its various components. We begin by reproducing the RecRanker pipeline through the implementation of all its key components. Our reproduction shows that the pairwise and listwise methods achieve a performance comparable to that reported in the original paper. For the pointwise method, while we are also able to reproduce the original paper's results, further analysis shows that the performance is abnormally high due to data leakage from the inclusion of ground-truth information in the prompts. To enable a fair and comprehensive evaluation of LLM-based top-k recommendations, we propose RecRankerEval, an extensible framework that covers five key dimensions: user sampling strategy, initial recommendation model, LLM backbone, dataset selection, and instruction tuning method. Using the RecRankerEval framework, we show that the original results of RecRanker can be reproduced on the ML-100K and ML-1M datasets, as well as the additional Amazon-Music dataset, but not on BookCrossing due to the lack of timestamp information in the original RecRanker paper. Furthermore, we demonstrate that RecRanker's performance can be improved by employing alternative user sampling methods, stronger initial recommenders, and more capable LLMs.
☆ On the Costs and Benefits of Learned Indexing for Dynamic High-Dimensional Data: Extended Version
One of the main challenges within the growing research area of learned indexing is the lack of adaptability to dynamically expanding datasets. This paper explores the dynamization of a static learned index for complex data through operations such as node splitting and broadening, enabling efficient adaptation to new data. Furthermore, we evaluate the trade-offs between static and dynamic approaches by introducing an amortized cost model to assess query performance in tandem with the build costs of the index structure, enabling experimental determination of when a dynamic learned index outperforms its static counterpart. We apply the dynamization method to a static learned index and demonstrate that its superior scaling quickly surpasses the static implementation in terms of overall costs as the database grows. This is an extended version of the paper presented at DAWAK 2025.
☆ KERAG_R: Knowledge-Enhanced Retrieval-Augmented Generation for Recommendation
Large Language Models (LLMs) have shown strong potential in recommender systems due to their contextual learning and generalisation capabilities. Existing LLM-based recommendation approaches typically formulate the recommendation task using specialised prompts designed to leverage their contextual abilities, and aligning their outputs closely with human preferences to yield an improved recommendation performance. However, the use of LLMs for recommendation tasks is limited by the absence of domain-specific knowledge. This lack of relevant relational knowledge about the items to be recommended in the LLM's pre-training corpus can lead to inaccuracies or hallucinations, resulting in incorrect or misleading recommendations. Moreover, directly using information from the knowledge graph introduces redundant and noisy information, which can affect the LLM's reasoning process or exceed its input context length, thereby reducing the performance of LLM-based recommendations. To address the lack of domain-specific knowledge, we propose a novel model called Knowledge-Enhanced Retrieval-Augmented Generation for Recommendation (KERAG_R). Specifically, we leverage a graph retrieval-augmented generation (GraphRAG) component to integrate additional information from a knowledge graph (KG) into instructions, enabling the LLM to collaboratively exploit recommendation signals from both text-based user interactions and the knowledge graph to better estimate the users' preferences in a recommendation context. In particular, we perform graph RAG by pre-training a graph attention network (GAT) to select the most relevant triple for the target users for the used LLM, thereby enhancing the LLM while reducing redundant and noisy information. Our extensive experiments on three public datasets show that our proposed KERAG_R model significantly outperforms ten existing state-of-the-art recommendation methods.
☆ Vers un cadre ontologique pour la gestion des comp{é}tences : {à} des fins de formation, de recrutement, de m{é}tier, ou de recherches associ{é}es
The rapid transformation of the labor market, driven by technological advancements and the digital economy, requires continuous competence development and constant adaptation. In this context, traditional competence management systems lack interoperability, adaptability, and semantic understanding, making it difficult to align individual competencies with labor market needs and training programs. This paper proposes an ontology-based framework for competence management, enabling a structured representation of competencies, occupations, and training programs. By leveraging ontological models and semantic reasoning, this framework aims to enhance the automation of competence-to-job matching, the personalization of learning recommendations, and career planning. This study discusses the design, implementation, and potential applications of the framework, focusing on competence training programs, job searching, and finding competent individuals.
comment: in French language. 36es Journ{\'e}es francophones d'Ing{\'e}nierie des Connaissances (IC 2025) @ Plate-Forme Intelligence Artificielle (PFIA 2025), Jul 2025, Dijon, France
☆ When Transformers Meet Recommenders: Integrating Self-Attentive Sequential Recommendation with Fine-Tuned LLMs
Self-Attentive Sequential Recommendation (SASRec) effectively captures long-term user preferences by applying attention mechanisms to historical interactions. Concurrently, the rise of Large Language Models (LLMs) has motivated research into LLM-based recommendation, which leverages their powerful generalization and language understanding capabilities. However, LLMs often lack the domain-specific knowledge and collaborative signals essential for high-quality recommendations when relying solely on textual prompts. To address this limitation, this study proposes SASRecLLM, a novel framework that integrates SASRec as a collaborative encoder with an LLM fine-tuned using Low-Rank Adaptation (LoRA). The components are connected via a mapping layer to align their dimensional spaces, and three targeted training strategies are designed to optimize the hybrid architecture. Extensive experiments on multiple datasets demonstrate that SASRecLLM achieves robust and consistent improvements over strong baselines in both cold-start and warm-start scenarios. This work advances the field of LLM-based recommendation by presenting a modular and effective paradigm for fusing structured collaborative filtering with the semantic power of fine-tuned LLMs. The implementation is available on GitHub: https://github.com/kechenkristin/RecLLM
☆ From ID-based to ID-free: Rethinking ID Effectiveness in Multimodal Collaborative Filtering Recommendation ACM MM'25
Most existing multimodal collaborative filtering recommendation (MCFRec) methods rely heavily on ID features and multimodal content to enhance recommendation performance. However, this paper reveals that ID features are effective but have limited benefits in multimodal collaborative filtering recommendation. Therefore, this paper systematically deconstruct the pros and cons of ID features: (i) they provide initial embedding but lack semantic richness, (ii) they provide a unique identifier for each user and item but hinder generalization to untrained data, and (iii) they assist in aligning and fusing multimodal features but may lead to representation shift. Based on these insights, this paper proposes IDFREE, an ID-free multimodal collaborative Filtering REcommEndation baseline. IDFREE replaces ID features with multimodal features and positional encodings to generate semantically meaningful ID-free embeddings. For ID-free multimodal collaborative filtering, it further proposes an adaptive similarity graph module to construct dynamic user-user and item-item graphs based on multimodal features. Then, an augmented user-item graph encoder is proposed to construct more effective user and item encoding. Finally, IDFREE achieves inter-multimodal alignment based on the contrastive learning and uses Softmax loss as recommendation loss. Basic experiments on three public datasets demonstrate that IDFREE outperforms existing ID-based MCFRec methods, achieving an average performance gain of 72.24% across standard metrics (Recall@5, 10, 20, 50 and NDCG@5, 10, 20, 50). Exploratory and extended experiments further validate our findings on the limitations of ID features in MCFRec. The code is released at https://github.com/G-H-Li/IDFREE.
comment: ACM MM'25 (Experimental supplementary version)
☆ SARA: Selective and Adaptive Retrieval-augmented Generation with Context Compression
Retrieval-augmented Generation (RAG) extends large language models (LLMs) with external knowledge but faces key challenges: restricted effective context length and redundancy in retrieved documents. Pure compression-based approaches reduce input size but often discard fine-grained details essential for factual accuracy. We propose SARA, a unified RAG framework that balances local precision and global knowledge coverage under tight context budgets. SARA combines natural-language text snippets with semantic compression vectors to jointly enhance context efficiency and answer correctness. It represents contexts at two complementary levels: 1) fine-grained natural-language spans that preserve critical entities and numerical values, and 2) compact, interpretable vectors that summarize high-level semantics. An iterative evidence-selection module employs the compression vectors for dynamic reranking of contexts. Across 9 datasets and 5 open-source LLMs spanning 3 model families (Mistral, Llama, and Gemma), SARA consistently improves answer relevance (+17.71), answer correctness (+13.72), and semantic similarity (+15.53), demonstrating the importance of integrating textual and compressed representations for robust, context-efficient RAG.
comment: 20 pages
☆ Beyond Retrieval: Ensembling Cross-Encoders and GPT Rerankers with LLMs for Biomedical QA
Biomedical semantic question answering rooted in information retrieval can play a crucial role in keeping up to date with vast, rapidly evolving and ever-growing biomedical literature. A robust system can help researchers, healthcare professionals and even layman users access relevant knowledge grounded in evidence. The BioASQ 2025 Task13b Challenge serves as an important benchmark, offering a competitive platform for advancement of this space. This paper presents the methodologies and results from our participation in this challenge where we built a Retrieval-Augmented Generation (RAG) system that can answer biomedical questions by retrieving relevant PubMed documents and snippets to generate answers. For the retrieval task, we generated dense embeddings from biomedical articles for initial retrieval, and applied an ensemble of finetuned cross-encoders and large language models (LLMs) for re-ranking to identify top relevant documents. Our solution achieved an MAP@10 of 0.1581, placing 10th on the leaderboard for the retrieval task. For answer generation, we employed few-shot prompting of instruction-tuned LLMs. Our system achieved macro-F1 score of 0.95 for yes/no questions (rank 12), Mean Reciprocal Rank (MRR) of 0.64 for factoid questions (rank 1), mean-F1 score of 0.63 for list questions (rank 5), and ROUGE-SU4 F1 score of 0.29 for ideal answers (rank 11).
comment: Paper submitted to CLEF 2025 CEUR-WS
♻ ☆ Uncertainty-Aware Complex Scientific Table Data Extraction
Table structure recognition (TSR) and optical character recognition (OCR) play crucial roles in extracting structured data from tables in scientific documents. However, existing extraction frameworks built on top of TSR and OCR methods often fail to quantify the uncertainties of extracted results. To obtain highly accurate data for scientific domains, all extracted data must be manually verified, which can be time-consuming and labor-intensive. We propose a framework that performs uncertainty-aware data extraction for complex scientific tables, built on conformal prediction, a model-agnostic method for uncertainty quantification (UQ). We explored various uncertainty scoring methods to aggregate the uncertainties introduced by TSR and OCR. We rigorously evaluated the framework using a standard benchmark and an in-house dataset consisting of complex scientific tables in six scientific domains. The results demonstrate the effectiveness of using UQ for extraction error detection, and by manually verifying only 47% of extraction results, the data quality can be improved by 30%. Our work quantitatively demonstrates the role of UQ with the potential of improving the efficiency in the human-machine cooperation process to obtain scientifically usable data from complex tables in scientific documents. All code and data are available on GitHub at https://github.com/lamps-lab/TSR-OCR-UQ/tree/main.
♻ ☆ Counterfactual Inference under Thompson Sampling RecSys '25
Recommender systems exemplify sequential decision-making under uncertainty, strategically deciding what content to serve to users, to optimise a range of potential objectives. To balance the explore-exploit trade-off successfully, Thompson sampling provides a natural and widespread paradigm to probabilistically select which action to take. Questions of causal and counterfactual inference, which underpin use-cases like offline evaluation, are not straightforward to answer in these contexts. Specifically, whilst most existing estimators rely on action propensities, these are not readily available under Thompson sampling procedures. We derive exact and efficiently computable expressions for action propensities under a variety of parameter and outcome distributions, enabling the use of off-policy estimators in Thompson sampling scenarios. This opens up a range of practical use-cases where counterfactual inference is crucial, including unbiased offline evaluation of recommender systems, as well as general applications of causal inference in online advertising, personalisation, and beyond.
comment: To appear in the Nineteenth ACM Conference on Recommender Systems (RecSys '25)
♻ ☆ PDFMathTranslate: Scientific Document Translation Preserving Layouts
Language barriers in scientific documents hinder the diffusion and development of science and technologies. However, prior efforts in translating such documents largely overlooked the information in layouts. To bridge the gap, we introduce PDFMathTranslate, the world's first open-source software for translating scientific documents while preserving layouts. Leveraging the most recent advances in large language models and precise layout detection, we contribute to the community with key improvements in precision, flexibility, and efficiency. The work has been open-sourced at https://github.com/byaidu/pdfmathtranslate with more than 222k downloads.
comment: 7 pages, 4 figures
♻ ☆ Multi-Channel Hypergraph Contrastive Learning for Matrix Completion
Rating is a typical user explicit feedback that visually reflects how much a user likes a related item. The (rating) matrix completion is essentially a rating prediction process, which is also a significant problem in recommender systems. Recently, graph neural networks (GNNs) have been widely used in matrix completion, which captures users' preferences over items by formulating a rating matrix as a bipartite graph. However, existing methods are susceptible due to data sparsity and long-tail distribution in real-world scenarios. Moreover, the messaging mechanism of GNNs makes it difficult to capture high-order correlations and constraints between nodes, which are essentially useful in recommendation tasks. To tackle these challenges, we propose a Multi-Channel Hypergraph Contrastive Learning framework for matrix completion, named MHCL. Specifically, MHCL adaptively learns hypergraph structures to capture high-order correlations between nodes and jointly captures local and global collaborative relationships through attention-based cross-view aggregation. Additionally, to consider the magnitude and order information of ratings, we treat different rating subgraphs as different channels, encourage alignment between adjacent ratings, and further achieve the mutual enhancement between different ratings through multi-channel cross-rating contrastive learning. Extensive experiments on five public datasets demonstrate that the proposed method significantly outperforms the current state-of-the-art approaches.
♻ ☆ RAG-R1 : Incentivize the Search and Reasoning Capabilities of LLMs through Multi-query Parallelism
Large Language Models (LLMs) have demonstrated remarkable capabilities across various tasks, while they remain prone to generating hallucinated or outdated responses due to their static internal knowledge. Recent advancements in Retrieval-Augmented Generation (RAG) methods have explored enhancing models' search and reasoning capabilities through reinforcement learning (RL). Although these methods demonstrate promising results, they face challenges in training stability and encounter issues such as substantial inference time and restricted capabilities due to the single-query mode. In this paper, we propose RAG-R1, a novel training framework designed to enable LLMs to adaptively leverage internal and external knowledge during the reasoning process. We further expand the generation and retrieval processes within the framework from single-query mode to multi-query parallelism, aimed at reducing inference time and enhancing the model's capabilities. Extensive experiments on seven question-answering benchmarks demonstrate that our method outperforms the strongest baseline by up to 13.2% and decreases inference time by 11.1%.
♻ ☆ SIGIR 2025 -- LiveRAG Challenge Report
The LiveRAG Challenge at SIGIR 2025, held between March and May 2025, provided a competitive platform for advancing Retrieval-Augmented Generation (RAG) technologies. Participants from academia and industry were invited to develop a RAG-based question-answering system using a fixed corpus (Fineweb-10BT) and a common open-source LLM (Falcon3-10B-Instruct). The goal was to facilitate challenging comparisons of retrieval and prompting strategies. During the Live Challenge Day, 70 teams from 27 different countries provided answers and supportive information to 500 unseen questions within a strict two-hour time window. Evaluation was conducted in two stages: first an automated LLM-as-a-judge approach was used to compute correctness and faithfulness score, then a manual review of top ranked submissions was conducted. The finalists were announced on June 12, 2025, with prizes awarded during the LiveRAG Workshop at SIGIR 2025 in Padua, Italy.
comment: 9 pages, 5 tables
♻ ☆ Do We Really Need Specialization? Evaluating Generalist Text Embeddings for Zero-Shot Recommendation and Search RecSys 2025
Pre-trained language models (PLMs) are widely used to derive semantic representations from item metadata in recommendation and search. In sequential recommendation, PLMs enhance ID-based embeddings through textual metadata, while in product search, they align item characteristics with user intent. Recent studies suggest task and domain-specific fine-tuning are needed to improve representational power. This paper challenges this assumption, showing that Generalist Text Embedding Models (GTEs), pre-trained on large-scale corpora, can guarantee strong zero-shot performance without specialized adaptation. Our experiments demonstrate that GTEs outperform traditional and fine-tuned models in both sequential recommendation and product search. We attribute this to a superior representational power, as they distribute features more evenly across the embedding space. Finally, we show that compressing embedding dimensions by focusing on the most informative directions (e.g., via PCA) effectively reduces noise and improves the performance of specialized models. To ensure reproducibility, we provide our repository at https://split.to/gte4ps.
comment: Accept as Short Paper at RecSys 2025
Multimedia 11
☆ Digital Wargames to Enhance Military Medical Evacuation Decision-Making
Medical evacuation is one of the United States Army's most storied and critical mission sets, responsible for efficiently and expediently evacuating the battlefield ill and injured. Medical evacuation planning involves designing a robust network of medical platforms and facilities capable of moving and treating large numbers of casualties. Until now, there has not been a medium to simulate these networks in a classroom setting and evaluate both offline planning and online decision-making performance. This work describes the Medical Evacuation Wargaming Initiative (MEWI), a three-dimensional multiplayer simulation developed in Unity that replicates battlefield constraints and uncertainties. MEWI accurately models patient interactions at casualty collection points, ambulance exchange points, medical treatment facilities, and evacuation platforms. Two operational scenarios are introduced: an amphibious island assault in the Pacific and a Eurasian conflict across a sprawling road and river network. These scenarios pit students against the clock to save as many casualties as possible while adhering to doctrinal lessons learned during didactic training. We visualize performance data collected from two iterations of the MEWI Pacific scenario executed in the United States Army's Medical Evacuation Doctrine Course. We consider post-wargame Likert survey data from student participants and external observer notes to identify key planning decision points, document medical evacuation lessons learned, and quantify general utility. Results indicate that MEWI participation substantially improves uptake of medical evacuation lessons learned and co-operative decision-making. MEWI is a substantial step forward in the field of high-fidelity training tools for medical education, and our study findings offer critical insights into improving medical evacuation education and operations across the joint force.
☆ MEDTalk: Multimodal Controlled 3D Facial Animation with Dynamic Emotions by Disentangled Embedding
Audio-driven emotional 3D facial animation aims to generate synchronized lip movements and vivid facial expressions. However, most existing approaches focus on static and predefined emotion labels, limiting their diversity and naturalness. To address these challenges, we propose MEDTalk, a novel framework for fine-grained and dynamic emotional talking head generation. Our approach first disentangles content and emotion embedding spaces from motion sequences using a carefully designed cross-reconstruction process, enabling independent control over lip movements and facial expressions. Beyond conventional audio-driven lip synchronization, we integrate audio and speech text, predicting frame-wise intensity variations and dynamically adjusting static emotion features to generate realistic emotional expressions. Furthermore, to enhance control and personalization, we incorporate multimodal inputs-including text descriptions and reference expression images-to guide the generation of user-specified facial expressions. With MetaHuman as the priority, our generated results can be conveniently integrated into the industrial production pipeline.
comment: 11 pages, 8 figures
☆ Remember Past, Anticipate Future: Learning Continual Multimodal Misinformation Detectors ACM MM 2025
Nowadays, misinformation articles, especially multimodal ones, are widely spread on social media platforms and cause serious negative effects. To control their propagation, Multimodal Misinformation Detection (MMD) becomes an active topic in the community to automatically identify misinformation. Previous MMD methods focus on supervising detectors by collecting offline data. However, in real-world scenarios, new events always continually emerge, making MMD models trained on offline data consistently outdated and ineffective. To address this issue, training MMD models under online data streams is an alternative, inducing an emerging task named continual MMD. Unfortunately, it is hindered by two major challenges. First, training on new data consistently decreases the detection performance on past data, named past knowledge forgetting. Second, the social environment constantly evolves over time, affecting the generalization on future data. To alleviate these challenges, we propose to remember past knowledge by isolating interference between event-specific parameters with a Dirichlet process-based mixture-of-expert structure, and anticipate future environmental distributions by learning a continuous-time dynamics model. Accordingly, we induce a new continual MMD method DAEDCMD. Extensive experiments demonstrate that DAEDCMD can consistently and significantly outperform the compared methods, including six MMD baselines and three continual learning methods.
comment: Accepted by ACM MM 2025. 10 pages, 6 figures. Code: https://github.com/wangbing1416/DAEDCMD
☆ D-FCGS: Feedforward Compression of Dynamic Gaussian Splatting for Free-Viewpoint Videos
Free-viewpoint video (FVV) enables immersive 3D experiences, but efficient compression of dynamic 3D representations remains a major challenge. Recent advances in 3D Gaussian Splatting (3DGS) and its dynamic extensions have enabled high-fidelity scene modeling. However, existing methods often couple scene reconstruction with optimization-dependent coding, which limits generalizability. This paper presents Feedforward Compression of Dynamic Gaussian Splatting (D-FCGS), a novel feedforward framework for compressing temporally correlated Gaussian point cloud sequences. Our approach introduces a Group-of-Frames (GoF) structure with I-P frame coding, where inter-frame motions are extracted via sparse control points. The resulting motion tensors are compressed in a feedforward manner using a dual prior-aware entropy model that combines hyperprior and spatial-temporal priors for accurate rate estimation. For reconstruction, we perform control-point-guided motion compensation and employ a refinement network to enhance view-consistent fidelity. Trained on multi-view video-derived Gaussian frames, D-FCGS generalizes across scenes without per-scene optimization. Experiments show that it matches the rate-distortion performance of optimization-based methods, achieving over 40 times compression in under 2 seconds while preserving visual quality across viewpoints. This work advances feedforward compression for dynamic 3DGS, paving the way for scalable FVV transmission and storage in immersive applications.
comment: 12 pages, 9 figures, 8 tables
☆ From ID-based to ID-free: Rethinking ID Effectiveness in Multimodal Collaborative Filtering Recommendation ACM MM'25
Most existing multimodal collaborative filtering recommendation (MCFRec) methods rely heavily on ID features and multimodal content to enhance recommendation performance. However, this paper reveals that ID features are effective but have limited benefits in multimodal collaborative filtering recommendation. Therefore, this paper systematically deconstruct the pros and cons of ID features: (i) they provide initial embedding but lack semantic richness, (ii) they provide a unique identifier for each user and item but hinder generalization to untrained data, and (iii) they assist in aligning and fusing multimodal features but may lead to representation shift. Based on these insights, this paper proposes IDFREE, an ID-free multimodal collaborative Filtering REcommEndation baseline. IDFREE replaces ID features with multimodal features and positional encodings to generate semantically meaningful ID-free embeddings. For ID-free multimodal collaborative filtering, it further proposes an adaptive similarity graph module to construct dynamic user-user and item-item graphs based on multimodal features. Then, an augmented user-item graph encoder is proposed to construct more effective user and item encoding. Finally, IDFREE achieves inter-multimodal alignment based on the contrastive learning and uses Softmax loss as recommendation loss. Basic experiments on three public datasets demonstrate that IDFREE outperforms existing ID-based MCFRec methods, achieving an average performance gain of 72.24% across standard metrics (Recall@5, 10, 20, 50 and NDCG@5, 10, 20, 50). Exploratory and extended experiments further validate our findings on the limitations of ID features in MCFRec. The code is released at https://github.com/G-H-Li/IDFREE.
comment: ACM MM'25 (Experimental supplementary version)
☆ AdaptaGen: Domain-Specific Image Generation through Hierarchical Semantic Optimization Framework
Domain-specific image generation aims to produce high-quality visual content for specialized fields while ensuring semantic accuracy and detail fidelity. However, existing methods exhibit two critical limitations: First, current approaches address prompt engineering and model adaptation separately, overlooking the inherent dependence between semantic understanding and visual representation in specialized domains. Second, these techniques inadequately incorporate domain-specific semantic constraints during content synthesis, resulting in generation outcomes that exhibit hallucinations and semantic deviations. To tackle these issues, we propose AdaptaGen, a hierarchical semantic optimization framework that integrates matrix-based prompt optimization with multi-perspective understanding, capturing comprehensive semantic relationships from both global and local perspectives. To mitigate hallucinations in specialized domains, we design a cross-modal adaptation mechanism, which, when combined with intelligent content synthesis, enables preserving core thematic elements while incorporating diverse details across images. Additionally, we introduce a two-phase caption semantic transformation during the generation phase. This approach maintains semantic coherence while enhancing visual diversity, ensuring the generated images adhere to domain-specific constraints. Experimental results confirm our approach's effectiveness, with our framework achieving superior performance across 40 categories from diverse datasets using only 16 images per category, demonstrating significant improvements in image quality, diversity, and semantic consistency.
♻ ☆ EEG2TEXT-CN: An Exploratory Study of Open-Vocabulary Chinese Text-EEG Alignment via Large Language Model and Contrastive Learning on ChineseEEG
We propose EEG2TEXT-CN, which, to the best of our knowledge, represents one of the earliest open-vocabulary EEG-to-text generation frameworks tailored for Chinese. Built on a biologically grounded EEG encoder (NICE-EEG) and a compact pretrained language model (MiniLM), our architecture aligns multichannel brain signals with natural language representations via masked pretraining and contrastive learning. Using a subset of the ChineseEEG dataset, where each sentence contains approximately ten Chinese characters aligned with 128-channel EEG recorded at 256 Hz, we segment EEG into per-character embeddings and predict full sentences in a zero-shot setting. The decoder is trained with teacher forcing and padding masks to accommodate variable-length sequences. Evaluation on over 1,500 training-validation sentences and 300 held-out test samples shows promising lexical alignment, with a best BLEU-1 score of 6.38\%. While syntactic fluency remains a challenge, our findings demonstrate the feasibility of non-phonetic, cross-modal language decoding from EEG. This work opens a new direction in multilingual brain-to-text research and lays the foundation for future cognitive-language interfaces in Chinese.
♻ ☆ What's Making That Sound Right Now? Video-centric Audio-Visual Localization ICCV 2025
Audio-Visual Localization (AVL) aims to identify sound-emitting sources within a visual scene. However, existing studies focus on image-level audio-visual associations, failing to capture temporal dynamics. Moreover, they assume simplified scenarios where sound sources are always visible and involve only a single object. To address these limitations, we propose AVATAR, a video-centric AVL benchmark that incorporates high-resolution temporal information. AVATAR introduces four distinct scenarios -- Single-sound, Mixed-sound, Multi-entity, and Off-screen -- enabling a more comprehensive evaluation of AVL models. Additionally, we present TAVLO, a novel video-centric AVL model that explicitly integrates temporal information. Experimental results show that conventional methods struggle to track temporal variations due to their reliance on global audio features and frame-level mappings. In contrast, TAVLO achieves robust and precise audio-visual alignment by leveraging high-resolution temporal modeling. Our work empirically demonstrates the importance of temporal dynamics in AVL and establishes a new standard for video-centric audio-visual localization.
comment: Published at ICCV 2025. Project page: https://hahyeon610.github.io/Video-centric_Audio_Visual_Localization/
♻ ☆ CFMW: Cross-modality Fusion Mamba for Robust Object Detection under Adverse Weather
Visible-infrared image pairs provide complementary information, enhancing the reliability and robustness of object detection applications in real-world scenarios. However, most existing methods face challenges in maintaining robustness under complex weather conditions, which limits their applicability. Meanwhile, the reliance on attention mechanisms in modality fusion introduces significant computational complexity and storage overhead, particularly when dealing with high-resolution images. To address these challenges, we propose the Cross-modality Fusion Mamba with Weather-removal (CFMW) to augment stability and cost-effectiveness under adverse weather conditions. Leveraging the proposed Perturbation-Adaptive Diffusion Model (PADM) and Cross-modality Fusion Mamba (CFM) modules, CFMW is able to reconstruct visual features affected by adverse weather, enriching the representation of image details. With efficient architecture design, CFMW is 3 times faster than Transformer-style fusion (e.g., CFT). To bridge the gap in relevant datasets, we construct a new Severe Weather Visible-Infrared (SWVI) dataset, encompassing diverse adverse weather scenarios such as rain, haze, and snow. The dataset contains 64,281 paired visible-infrared images, providing a valuable resource for future research. Extensive experiments on public datasets (i.e., M3FD and LLVIP) and the newly constructed SWVI dataset conclusively demonstrate that CFMW achieves state-of-the-art detection performance. Both the dataset and source code will be made publicly available at https://github.com/lhy-zjut/CFMW.
comment: Accepted to IEEE Transactions on Circuits and Systems for Video Technology (TCSVT). The dataset and source code will be made publicly available at https://github.com/lhy-zjut/CFMW
♻ ☆ UniCode$^2$: Cascaded Large-scale Codebooks for Unified Multimodal Understanding and Generation
Unified multimodal large language models (MLLMs) have shown promise in jointly advancing multimodal understanding and generation, with visual codebooks discretizing images into tokens for autoregressive modeling. Existing codebook-based methods either rely on small vocabularies (~16K entries) that lack fine-grained semantics or naively scale up, resulting in low token utilization and unstable training. We propose UniCode$^2$, a cascaded codebook framework enabling large-scale, semantically aligned, and stable visual tokenization. By clustering millions of SigLIP sequence embeddings, we build a 500K-entry codebook that preserves vision-language alignment while expanding capacity. Stability is ensured via a cascaded design: a frozen codebook anchors the embedding space, and a trainable codebook refines task-specific semantics. This decoupling promotes high utilization and robust learning. Moreover, the alignment of our visual tokens with textual semantics enables seamless integration with pretrained diffusion decoders, supporting high-quality visual synthesis with minimal adaptation. UniCode^2 delivers strong performance across diverse benchmarks, demonstrating the viability of scaling visual token spaces without sacrificing stability, semantics, or modularity.
comment: 19 pages, 5 figures
♻ ☆ Seeing Sarcasm Through Different Eyes: Analyzing Multimodal Sarcasm Perception in Large Vision-Language Models
With the advent of large vision-language models (LVLMs) demonstrating increasingly human-like abilities, a pivotal question emerges: do different LVLMs interpret multimodal sarcasm differently, and can a single model grasp sarcasm from multiple perspectives like humans? To explore this, we introduce an analytical framework using systematically designed prompts on existing multimodal sarcasm datasets. Evaluating 12 state-of-the-art LVLMs over 2,409 samples, we examine interpretive variations within and across models, focusing on confidence levels, alignment with dataset labels, and recognition of ambiguous "neutral" cases. Our findings reveal notable discrepancies -- across LVLMs and within the same model under varied prompts. While classification-oriented prompts yield higher internal consistency, models diverge markedly when tasked with interpretive reasoning. These results challenge binary labeling paradigms by highlighting sarcasm's subjectivity. We advocate moving beyond rigid annotation schemes toward multi-perspective, uncertainty-aware modeling, offering deeper insights into multimodal sarcasm comprehension. Our code and data are available at: https://github.com/CoderChen01/LVLMSarcasmAnalysis
Information Retrieval 18
☆ Information Needs and Practices Supported by ChatGPT
This study considers ChatGPT as an information source, investigating the information needs that people come to ChatGPT with and the information practices that ChatGPT supports, through a qualitative content analysis of 205 user vignettes. The findings show that ChatGPT is used in a range of life domains (home/family, work, leisure, etc.) and for a range of human needs (writing/editing, learning, simple programming tasks, etc.), constituting the information needs that people use ChatGPT to address. Related to these information needs, the findings show six categories of information practices that ChatGPT supports: Writing, Deciding, Identifying, Ideating, Talking, and Critiquing. This work suggests that, in the AI age, information need should be conceptualized not just as a matter of "getting questions answered" or even "making sense," but as skillfully coping in the world, a notion that includes both understanding and action. This study leads to numerous opportunities for future work at the junction of generative AI and information needs, seeking, use and experience.
comment: To be presented at the 2025 ASIS&T virtual satellite meeting, December 2025
☆ In-Context Learning as an Effective Estimator of Functional Correctness of LLM-Generated Code
When applying LLM-based code generation to software development projects that follow a feature-driven or rapid application development approach, it becomes necessary to estimate the functional correctness of the generated code in the absence of test cases. Just as a user selects a relevant document from a ranked list of retrieved ones, a software generation workflow requires a developer to choose (and potentially refine) a generated solution from a ranked list of alternative solutions, ordered by their posterior likelihoods. This implies that estimating the quality of a ranked list -- akin to estimating "relevance" for query performance prediction (QPP) in IR -- is also crucial for generative software development, where quality is defined in terms of "functional correctness". In this paper, we propose an in-context learning (ICL) based approach for code quality estimation. Our findings demonstrate that providing few-shot examples of functionally correct code from a training set enhances the performance of existing QPP approaches as well as a zero-shot-based approach for code quality estimation.
☆ Interest Networks (iNETs) for Cities: Cross-Platform Insights and Urban Behavior Explanations KDD
Location-Based Social Networks (LBSNs) provide a rich foundation for modeling urban behavior through iNETs (Interest Networks), which capture how user interests are distributed throughout urban spaces. This study compares iNETs across platforms (Google Places and Foursquare) and spatial granularities, showing that coarser levels reveal more consistent cross-platform patterns, while finer granularities expose subtle, platform-specific behaviors. Our analysis finds that, in general, user interest is primarily shaped by geographic proximity and venue similarity, while socioeconomic and political contexts play a lesser role. Building on these insights, we develop a multi-level, explainable recommendation system that predicts high-interest urban regions for different user types. The model adapts to behavior profiles -- such as explorers, who are driven by proximity, and returners, who prefer familiar venues -- and provides natural-language explanations using explainable AI (XAI) techniques. To support our approach, we introduce h3-cities, a tool for multi-scale spatial analysis, and release a public demo for interactively exploring personalized urban recommendations. Our findings contribute to urban mobility research by providing scalable, context-aware, and interpretable recommendation systems.
comment: Accepted at ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD-UMC)
☆ SimLab: A Platform for Simulation-based Evaluation of Conversational Information Access Systems
Research on interactive and conversational information access systems, including search engines, recommender systems, and conversational assistants, has been hindered by the difficulty in evaluating such systems with reproducible experiments. User simulation provides a promising solution, but there is a lack of infrastructure and tooling to support this kind of evaluation. To facilitate simulation-based evaluation of conversational information access systems, we introduce SimLab, the first cloud-based platform to provide a centralized general solution for the community to benchmark both conversational systems and user simulators in a controlled and reproducible environment. We articulate requirements for such a platform and propose a general infrastructure to address these requirements. We then present the design and implementation of an initial version of SimLab and showcase its features with an initial evaluation task of conversational movie recommendation, which is made publicly available. Furthermore, we discuss the sustainability of the platform and its future opportunities. This paper is a call for the community to contribute to the platform to drive progress in the field of conversational information access and user simulation.
☆ PLACE: Prompt Learning for Attributed Community Search
In this paper, we propose PLACE (Prompt Learning for Attributed Community Search), an innovative graph prompt learning framework for ACS. Enlightened by prompt-tuning in Natural Language Processing (NLP), where learnable prompt tokens are inserted to contextualize NLP queries, PLACE integrates structural and learnable prompt tokens into the graph as a query-dependent refinement mechanism, forming a prompt-augmented graph. Within this prompt-augmented graph structure, the learned prompt tokens serve as a bridge that strengthens connections between graph nodes for the query, enabling the GNN to more effectively identify patterns of structural cohesiveness and attribute similarity related to the specific query. We employ an alternating training paradigm to optimize both the prompt parameters and the GNN jointly. Moreover, we design a divide-and-conquer strategy to enhance scalability, supporting the model to handle million-scale graphs. Extensive experiments on 9 real-world graphs demonstrate the effectiveness of PLACE for three types of ACS queries, where PLACE achieves higher F1 scores by 22% compared to the state-of-the-arts on average.
comment: 15 pages, 9 figures
☆ Phantom Subgroup Poisoning: Stealth Attacks on Federated Recommender Systems
Federated recommender systems (FedRec) have emerged as a promising solution for delivering personalized recommendations while safeguarding user privacy. However, recent studies have demonstrated their vulnerability to poisoning attacks. Existing attacks typically target the entire user group, which compromises stealth and increases the risk of detection. In contrast, real-world adversaries may prefer to prompt target items to specific user subgroups, such as recommending health supplements to elderly users. Motivated by this gap, we introduce Spattack, the first targeted poisoning attack designed to manipulate recommendations for specific user subgroups in the federated setting. Specifically, Spattack adopts a two-stage approximation-and-promotion strategy, which first simulates user embeddings of target/non-target subgroups and then prompts target items to the target subgroups. To enhance the approximation stage, we push the inter-group embeddings away based on contrastive learning and augment the target group's relevant item set based on clustering. To enhance the promotion stage, we further propose to adaptively tune the optimization weights between target and non-target subgroups. Besides, an embedding alignment strategy is proposed to align the embeddings between the target items and the relevant items. We conduct comprehensive experiments on three real-world datasets, comparing Spattack against seven state-of-the-art poisoning attacks and seven representative defense mechanisms. Experimental results demonstrate that Spattack consistently achieves strong manipulation performance on the specific user subgroup, while incurring minimal impact on non-target users, even when only 0.1\% of users are malicious. Moreover, Spattack maintains competitive overall recommendation performance and exhibits strong resilience against existing mainstream defenses.
comment: 13 pages
☆ Harnessing Pairwise Ranking Prompting Through Sample-Efficient Ranking Distillation SIGIR 2025
While Pairwise Ranking Prompting (PRP) with Large Language Models (LLMs) is one of the most effective zero-shot document ranking methods, it has a quadratic computational complexity with respect to the number of documents to be ranked, as it requires an enumeration over all possible document pairs. Consequently, the outstanding ranking performance of PRP has remained unreachable for most real-world ranking applications. In this work, we propose to harness the effectiveness of PRP through pairwise distillation. Specifically, we distill a pointwise student ranker from pairwise teacher labels generated by PRP, resulting in an efficient student model that retains the performance of PRP with substantially lower computational costs. Furthermore, we find that the distillation process can be made sample-efficient: with only 2% of pairs, we are able to obtain the same performance as using all pairs for teacher labels. Thus, our novel approach provides a solution to harness the ranking performance of PRP without incurring high computational costs during both distillation and serving.
comment: ReNeuIR 2025 (at SIGIR 2025) - 4th Workshop on Reaching Efficiency in Neural Information Retrieval, July 17, 2025, Padua, Italy
☆ "This Suits You the Best": Query Focused Comparative Explainable Summarization
Product recommendations inherently involve comparisons, yet traditional opinion summarization often fails to provide holistic comparative insights. We propose the novel task of generating Query-Focused Comparative Explainable Summaries (QF-CES) using Multi-Source Opinion Summarization (M-OS). To address the lack of query-focused recommendation datasets, we introduce MS-Q2P, comprising 7,500 queries mapped to 22,500 recommended products with metadata. We leverage Large Language Models (LLMs) to generate tabular comparative summaries with query-specific explanations. Our approach is personalized, privacy-preserving, recommendation engine-agnostic, and category-agnostic. M-OS as an intermediate step reduces inference latency approximately by 40% compared to the direct input approach (DIA), which processes raw data directly. We evaluate open-source and proprietary LLMs for generating and assessing QF-CES. Extensive evaluations using QF-CES-PROMPT across 5 dimensions (clarity, faithfulness, informativeness, format adherence, and query relevance) showed an average Spearman correlation of 0.74 with human judgments, indicating its potential for QF-CES evaluation.
☆ FindRec: Stein-Guided Entropic Flow for Multi-Modal Sequential Recommendation KDD 2025
Modern recommendation systems face significant challenges in processing multimodal sequential data, particularly in temporal dynamics modeling and information flow coordination. Traditional approaches struggle with distribution discrepancies between heterogeneous features and noise interference in multimodal signals. We propose \textbf{FindRec}~ (\textbf{F}lexible unified \textbf{in}formation \textbf{d}isentanglement for multi-modal sequential \textbf{Rec}ommendation), introducing a novel "information flow-control-output" paradigm. The framework features two key innovations: (1) A Stein kernel-based Integrated Information Coordination Module (IICM) that theoretically guarantees distribution consistency between multimodal features and ID streams, and (2) A cross-modal expert routing mechanism that adaptively filters and combines multimodal features based on their contextual relevance. Our approach leverages multi-head subspace decomposition for routing stability and RBF-Stein gradient for unbiased distribution alignment, enhanced by linear-complexity Mamba layers for efficient temporal modeling. Extensive experiments on three real-world datasets demonstrate FindRec's superior performance over state-of-the-art baselines, particularly in handling long sequences and noisy multimodal inputs. Our framework achieves both improved recommendation accuracy and enhanced model interpretability through its modular design. The implementation code is available anonymously online for easy reproducibility~\footnote{https://github.com/Applied-Machine-Learning-Lab/FindRec}.
comment: Accepted by KDD 2025
☆ Heterogeneous User Modeling for LLM-based Recommendation RecSys 2025
Leveraging Large Language Models (LLMs) for recommendation has demonstrated notable success in various domains, showcasing their potential for open-domain recommendation. A key challenge to advancing open-domain recommendation lies in effectively modeling user preferences from users' heterogeneous behaviors across multiple domains. Existing approaches, including ID-based and semantic-based modeling, struggle with poor generalization, an inability to compress noisy interactions effectively, and the domain seesaw phenomenon. To address these challenges, we propose a Heterogeneous User Modeling (HUM) method, which incorporates a compression enhancer and a robustness enhancer for LLM-based recommendation. The compression enhancer uses a customized prompt to compress heterogeneous behaviors into a tailored token, while a masking mechanism enhances cross-domain knowledge extraction and understanding. The robustness enhancer introduces a domain importance score to mitigate the domain seesaw phenomenon by guiding domain optimization. Extensive experiments on heterogeneous datasets validate that HUM effectively models user heterogeneity by achieving both high efficacy and robustness, leading to superior performance in open-domain recommendation.
comment: Accepted by RecSys 2025
☆ Hierarchical Intent-guided Optimization with Pluggable LLM-Driven Semantics for Session-based Recommendation
Session-based Recommendation (SBR) aims to predict the next item a user will likely engage with, using their interaction sequence within an anonymous session. Existing SBR models often focus only on single-session information, ignoring inter-session relationships and valuable cross-session insights. Some methods try to include inter-session data but struggle with noise and irrelevant information, reducing performance. Additionally, most models rely on item ID co-occurrence and overlook rich semantic details, limiting their ability to capture fine-grained item features. To address these challenges, we propose a novel hierarchical intent-guided optimization approach with pluggable LLM-driven semantic learning for session-based recommendations, called HIPHOP. First, we introduce a pluggable embedding module based on large language models (LLMs) to generate high-quality semantic representations, enhancing item embeddings. Second, HIPHOP utilizes graph neural networks (GNNs) to model item transition relationships and incorporates a dynamic multi-intent capturing module to address users' diverse interests within a session. Additionally, we design a hierarchical inter-session similarity learning module, guided by user intent, to capture global and local session relationships, effectively exploring users' long-term and short-term interests. To mitigate noise, an intent-guided denoising strategy is applied during inter-session learning. Finally, we enhance the model's discriminative capability by using contrastive learning to optimize session representations. Experiments on multiple datasets show that HIPHOP significantly outperforms existing methods, demonstrating its effectiveness in improving recommendation quality. Our code is available: https://github.com/hjx159/HIPHOP.
☆ News Source Citing Patterns in AI Search Systems
AI-powered search systems are emerging as new information gatekeepers, fundamentally transforming how users access news and information. Despite their growing influence, the citation patterns of these systems remain poorly understood. We address this gap by analyzing data from the AI Search Arena, a head-to-head evaluation platform for AI search systems. The dataset comprises over 24,000 conversations and 65,000 responses from models across three major providers: OpenAI, Perplexity, and Google. Among the over 366,000 citations embedded in these responses, 9% reference news sources. We find that while models from different providers cite distinct news sources, they exhibit shared patterns in citation behavior. News citations concentrate heavily among a small number of outlets and display a pronounced liberal bias, though low-credibility sources are rarely cited. User preference analysis reveals that neither the political leaning nor the quality of cited news sources significantly influences user satisfaction. These findings reveal significant challenges in current AI search systems and have important implications for their design and governance.
comment: 15 pages, 7 figures
♻ ☆ Late Chunking: Contextual Chunk Embeddings Using Long-Context Embedding Models
Many use cases require retrieving smaller portions of text, and dense vector-based retrieval systems often perform better with shorter text segments, as the semantics are less likely to be over-compressed in the embeddings. Consequently, practitioners often split text documents into smaller chunks and encode them separately. However, chunk embeddings created in this way can lose contextual information from surrounding chunks, resulting in sub-optimal representations. In this paper, we introduce a novel method called late chunking, which leverages long context embedding models to first embed all tokens of the long text, with chunking applied after the transformer model and just before mean pooling - hence the term late in its naming. The resulting chunk embeddings capture the full contextual information, leading to superior results across various retrieval tasks. The method is generic enough to be applied to a wide range of long-context embedding models and works without additional training. To further increase the effectiveness of late chunking, we propose a dedicated fine-tuning approach for embedding models.
comment: 11 pages, 3rd draft
♻ ☆ jina-embeddings-v4: Universal Embeddings for Multimodal Multilingual Retrieval
We introduce jina-embeddings-v4, a 3.8 billion parameter multimodal embedding model that unifies text and image representations through a novel architecture supporting both single-vector and multi-vector embeddings in the late interaction style. The model incorporates task-specific Low-Rank Adaptation (LoRA) adapters to optimize performance across diverse retrieval scenarios, including query-document retrieval, semantic text similarity, and code search. Comprehensive evaluations demonstrate that jina-embeddings-v4 achieves state-of-the-art performance on both single-modal and cross-modal retrieval tasks, with particular strength in processing visually rich content such as tables, charts, diagrams, and mixed-media formats. To facilitate evaluation of this capability, we also introduce Jina-VDR, a novel benchmark specifically designed for visually rich image retrieval.
comment: 22 pages, 1-10 main, 14-22 experimental results, benchmark tables
♻ ☆ A Generalised and Adaptable Reinforcement Learning Stopping Method SIGIR2025
This paper presents a Technology Assisted Review (TAR) stopping approach based on Reinforcement Learning (RL). Previous such approaches offered limited control over stopping behaviour, such as fixing the target recall and tradeoff between preferring to maximise recall or cost. These limitations are overcome by introducing a novel RL environment, GRLStop, that allows a single model to be applied to multiple target recalls, balances the recall/cost tradeoff and integrates a classifier. Experiments were carried out on six benchmark datasets (CLEF e-Health datasets 2017-9, TREC Total Recall, TREC Legal and Reuters RCV1) at multiple target recall levels. Results showed that the proposed approach to be effective compared to multiple baselines in addition to offering greater flexibility.
comment: Accepted by SIGIR2025, Figure 4 legend updated
♻ ☆ Why Do Open-Source LLMs Struggle with Data Analysis? A Systematic Empirical Study
Large Language Models (LLMs) hold promise in automating data analysis tasks, yet open-source models face significant limitations in these kinds of reasoning-intensive scenarios. In this work, we investigate strategies to enhance the data analysis capabilities of open-source LLMs. By curating a seed dataset of diverse, realistic scenarios, we evaluate models across three dimensions: data understanding, code generation, and strategic planning. Our analysis reveals three key findings: (1) Strategic planning quality serves as the primary determinant of model performance; (2) Interaction design and task complexity significantly influence reasoning capabilities; (3) Data quality demonstrates a greater impact than diversity in achieving optimal performance. We leverage these insights to develop a data synthesis methodology, demonstrating significant improvements in open-source LLMs' analytical reasoning capabilities.
comment: Work in progress
♻ ☆ Are Information Retrieval Approaches Good at Harmonising Longitudinal Survey Questions in Social Science? SIGIR 2025
Automated detection of semantically equivalent questions in longitudinal social science surveys is crucial for long-term studies informing empirical research in the social, economic, and health sciences. Retrieving equivalent questions faces dual challenges: inconsistent representation of theoretical constructs (i.e. concept/sub-concept) across studies as well as between question and response options, and the evolution of vocabulary and structure in longitudinal text. To address these challenges, our multi-disciplinary collaboration of computer scientists and survey specialists presents a new information retrieval (IR) task of identifying concept (e.g. Housing, Job, etc.) equivalence across question and response options to harmonise longitudinal population studies. This paper investigates multiple unsupervised approaches on a survey dataset spanning 1946-2020, including probabilistic models, linear probing of language models, and pre-trained neural networks specialised for IR. We show that IR-specialised neural models achieve the highest overall performance with other approaches performing comparably. Additionally, the re-ranking of the probabilistic model's results with neural models only introduces modest improvements of 0.07 at most in F1-score. Qualitative post-hoc evaluation by survey specialists shows that models generally have a low sensitivity to questions with high lexical overlap, particularly in cases where sub-concepts are mismatched. Altogether, our analysis serves to further research on harmonising longitudinal studies in social science.
comment: Accepted at SIGIR 2025
♻ ☆ A Text is Worth Several Tokens: Text Embedding from LLMs Secretly Aligns Well with The Key Tokens ACL2025
Text embeddings from large language models (LLMs) have achieved excellent results in tasks such as information retrieval, semantic textual similarity, etc. In this work, we show an interesting finding: when feeding a text into the LLM-based embedder, the obtained text embedding will be able to be aligned with the key tokens in the input text. We first fully analyze this phenomenon on eight LLM-based embedders and show that this phenomenon is universal and is not affected by model architecture, training strategy, and embedding method. With a deeper analysis, we find that the main change in embedding space between these embedders and their LLM backbones is in the first principal component. By adjusting the first principal component, we can align text embedding with the key tokens. Finally, we give several examples to demonstrate the vast application potential of this finding: (1) we propose a simple and practical sparse retrieval method based on the aligned tokens, which can achieve 80% of the dense retrieval effect of the same model while reducing the computation significantly; (2) we show that our findings provide a novel perspective to help understand novel technologies (e.g., instruction-following embedding) and fuzzy concepts (e.g., semantic relatedness vs. similarity) in this field.
comment: ACL2025 Oral
Multimedia 10
☆ NavigScene: Bridging Local Perception and Global Navigation for Beyond-Visual-Range Autonomous Driving
Autonomous driving systems have made significant advances in Q&A, perception, prediction, and planning based on local visual information, yet they struggle to incorporate broader navigational context that human drivers routinely utilize. We address this critical gap between local sensor data and global navigation information by proposing NavigScene, an auxiliary navigation-guided natural language dataset that simulates a human-like driving environment within autonomous driving systems. Moreover, we develop three complementary paradigms to leverage NavigScene: (1) Navigation-guided Reasoning, which enhances vision-language models by incorporating navigation context into the prompting approach; (2) Navigation-guided Preference Optimization, a reinforcement learning method that extends Direct Preference Optimization to improve vision-language model responses by establishing preferences for navigation-relevant summarized information; and (3) Navigation-guided Vision-Language-Action model, which integrates navigation guidance and vision-language models with conventional driving models through feature fusion. Extensive experiments demonstrate that our approaches significantly improve performance across perception, prediction, planning, and question-answering tasks by enabling reasoning capabilities beyond visual range and improving generalization to diverse driving scenarios. This work represents a significant step toward more comprehensive autonomous driving systems capable of navigating complex, unfamiliar environments with greater reliability and safety.
comment: Accepted by ACM Multimedia 2025
☆ CLIP-Guided Backdoor Defense through Entropy-Based Poisoned Dataset Separation
Deep Neural Networks (DNNs) are susceptible to backdoor attacks, where adversaries poison training data to implant backdoor into the victim model. Current backdoor defenses on poisoned data often suffer from high computational costs or low effectiveness against advanced attacks like clean-label and clean-image backdoors. To address them, we introduce CLIP-Guided backdoor Defense (CGD), an efficient and effective method that mitigates various backdoor attacks. CGD utilizes a publicly accessible CLIP model to identify inputs that are likely to be clean or poisoned. It then retrains the model with these inputs, using CLIP's logits as a guidance to effectively neutralize the backdoor. Experiments on 4 datasets and 11 attack types demonstrate that CGD reduces attack success rates (ASRs) to below 1% while maintaining clean accuracy (CA) with a maximum drop of only 0.3%, outperforming existing defenses. Additionally, we show that clean-data-based defenses can be adapted to poisoned data using CGD. Also, CGD exhibits strong robustness, maintaining low ASRs even when employing a weaker CLIP model or when CLIP itself is compromised by a backdoor. These findings underscore CGD's exceptional efficiency, effectiveness, and applicability for real-world backdoor defense scenarios. Code: https://github.com/binyxu/CGD.
comment: 15 pages, 9 figures, 15 tables. To appear in the Proceedings of the 32nd ACM International Conference on Multimedia (MM '25)
☆ Hear-Your-Click: Interactive Video-to-Audio Generation via Object-aware Contrastive Audio-Visual Fine-tuning
Video-to-audio (V2A) generation shows great potential in fields such as film production. Despite significant advances, current V2A methods, which rely on global video information, struggle with complex scenes and often fail to generate audio tailored to specific objects or regions in the videos. To address these limitations, we introduce Hear-Your-Click, an interactive V2A framework that enables users to generate sounds for specific objects in the videos by simply clicking on the frame. To achieve this, we propose Object-aware Contrastive Audio-Visual Fine-tuning (OCAV) with a Mask-guided Visual Encoder (MVE) to obtain object-level visual features aligned with corresponding audio segments. Furthermore, we tailor two data augmentation strategies: Random Video Stitching (RVS) and Mask-guided Loudness Modulation (MLM), aimed at enhancing the model's sensitivity to the segmented objects. To effectively measure the audio-visual correspondence, we design a new evaluation metric, the CAV score, for evaluation. Extensive experiments demonstrate that our framework offers more precise control and improved generation performance across various metrics. Project Page: https://github.com/SynapGrid/Hear-Your-Click
☆ Boosting Temporal Sentence Grounding via Causal Inference ACM MM 2025
Temporal Sentence Grounding (TSG) aims to identify relevant moments in an untrimmed video that semantically correspond to a given textual query. Despite existing studies having made substantial progress, they often overlook the issue of spurious correlations between video and textual queries. These spurious correlations arise from two primary factors: (1) inherent biases in the textual data, such as frequent co-occurrences of specific verbs or phrases, and (2) the model's tendency to overfit to salient or repetitive patterns in video content. Such biases mislead the model into associating textual cues with incorrect visual moments, resulting in unreliable predictions and poor generalization to out-of-distribution examples. To overcome these limitations, we propose a novel TSG framework, causal intervention and counterfactual reasoning that utilizes causal inference to eliminate spurious correlations and enhance the model's robustness. Specifically, we first formulate the TSG task from a causal perspective with a structural causal model. Then, to address unobserved confounders reflecting textual biases toward specific verbs or phrases, a textual causal intervention is proposed, utilizing do-calculus to estimate the causal effects. Furthermore, visual counterfactual reasoning is performed by constructing a counterfactual scenario that focuses solely on video features, excluding the query and fused multi-modal features. This allows us to debias the model by isolating and removing the influence of the video from the overall effect. Experiments on public datasets demonstrate the superiority of the proposed method. The code is available at https://github.com/Tangkfan/CICR.
comment: Accepted by ACM MM 2025
☆ EXPOTION: Facial Expression and Motion Control for Multimodal Music Generation
We propose Expotion (Facial Expression and Motion Control for Multimodal Music Generation), a generative model leveraging multimodal visual controls - specifically, human facial expressions and upper-body motion - as well as text prompts to produce expressive and temporally accurate music. We adopt parameter-efficient fine-tuning (PEFT) on the pretrained text-to-music generation model, enabling fine-grained adaptation to the multimodal controls using a small dataset. To ensure precise synchronization between video and music, we introduce a temporal smoothing strategy to align multiple modalities. Experiments demonstrate that integrating visual features alongside textual descriptions enhances the overall quality of generated music in terms of musicality, creativity, beat-tempo consistency, temporal alignment with the video, and text adherence, surpassing both proposed baselines and existing state-of-the-art video-to-music generation models. Additionally, we introduce a novel dataset consisting of 7 hours of synchronized video recordings capturing expressive facial and upper-body gestures aligned with corresponding music, providing significant potential for future research in multimodal and interactive music generation.
☆ Improving BERT for Symbolic Music Understanding Using Token Denoising and Pianoroll Prediction
We propose a pre-trained BERT-like model for symbolic music understanding that achieves competitive performance across a wide range of downstream tasks. To achieve this target, we design two novel pre-training objectives, namely token correction and pianoroll prediction. First, we sample a portion of note tokens and corrupt them with a limited amount of noise, and then train the model to denoise the corrupted tokens; second, we also train the model to predict bar-level and local pianoroll-derived representations from the corrupted note tokens. We argue that these objectives guide the model to better learn specific musical knowledge such as pitch intervals. For evaluation, we propose a benchmark that incorporates 12 downstream tasks ranging from chord estimation to symbolic genre classification. Results confirm the effectiveness of the proposed pre-training objectives on downstream tasks.
comment: Accepted at ISMIR 2025
☆ Music2Palette: Emotion-aligned Color Palette Generation via Cross-Modal Representation Learning
Emotion alignment between music and palettes is crucial for effective multimedia content, yet misalignment creates confusion that weakens the intended message. However, existing methods often generate only a single dominant color, missing emotion variation. Others rely on indirect mappings through text or images, resulting in the loss of crucial emotion details. To address these challenges, we present Music2Palette, a novel method for emotion-aligned color palette generation via cross-modal representation learning. We first construct MuCED, a dataset of 2,634 expert-validated music-palette pairs aligned through Russell-based emotion vectors. To directly translate music into palettes, we propose a cross-modal representation learning framework with a music encoder and color decoder. We further propose a multi-objective optimization approach that jointly enhances emotion alignment, color diversity, and palette coherence. Extensive experiments demonstrate that our method outperforms current methods in interpreting music emotion and generating attractive and diverse color palettes. Our approach enables applications like music-driven image recoloring, video generating, and data visualization, bridging the gap between auditory and visual emotion experiences.
♻ ☆ SVLA: A Unified Speech-Vision-Language Assistant with Multimodal Reasoning and Speech Generation
Large vision and language models show strong performance in tasks like image captioning, visual question answering, and retrieval. However, challenges remain in integrating speech, text, and vision into a unified model, especially for spoken tasks. Speech generation methods vary (some produce speech directly), others through text (but their impact on quality is unclear). Evaluation often relies on automatic speech recognition, which may introduce bias. We propose SVLA, a unified speech vision language model based on a transformer architecture that handles multimodal inputs and outputs. We train it on 38.2 million speech text image examples, including 64.1 hours of synthetic speech. We also introduce Speech VQA Accuracy, a new metric for evaluating spoken responses. SVLA improves multimodal understanding and generation by better combining speech, vision, and language.
comment: 21 pages
♻ ☆ UniForm: A Unified Multi-Task Diffusion Transformer for Audio-Video Generation
With the rise of diffusion models, audio-video generation has been revolutionized. However, most existing methods rely on separate modules for each modality, with limited exploration of unified generative architectures. In addition, many are confined to a single task and small-scale datasets. To overcome these limitations, we introduce UniForm, a unified multi-task diffusion transformer that generates both audio and visual modalities in a shared latent space. By using a unified denoising network, UniForm captures the inherent correlations between sound and vision. Additionally, we propose task-specific noise schemes and task tokens, enabling the model to support multiple tasks with a single set of parameters, including video-to-audio, audio-to-video and text-to-audio-video generation. Furthermore, by leveraging large language models and a large-scale text-audio-video combined dataset, UniForm achieves greater generative diversity than prior approaches. Experiments show that UniForm achieves performance close to the state-of-the-art single-task models across three generation tasks, with generated content that is not only highly aligned with real-world data distributions but also enables more diverse and fine-grained generation.
comment: Our demos are available at https://uniform-t2av.github.io/
♻ ☆ Manipulated Regions Localization For Partially Deepfake Audio: A Survey
With the development of audio deepfake techniques, attacks with partially deepfake audio are beginning to rise. Compared to fully deepfake, it is much harder to be identified by the detector due to the partially cryptic manipulation, resulting in higher security risks. Although some studies have been launched, there is no comprehensive review to systematically introduce the current situations and development trends for addressing this issue. Thus, in this survey, we are the first to outline a systematic introduction for partially deepfake audio manipulated region localization tasks, including the fundamentals, branches of existing methods, current limitations and potential trends, providing a revealing insight into this scope.
comment: Submitted to IEEE Transactions on Pattern Analysis and Machine Intelligence
Information Retrieval 6
☆ Multimedia Verification Through Multi-Agent Deep Research Multimodal Large Language Models
This paper presents our submission to the ACMMM25 - Grand Challenge on Multimedia Verification. We developed a multi-agent verification system that combines Multimodal Large Language Models (MLLMs) with specialized verification tools to detect multimedia misinformation. Our system operates through six stages: raw data processing, planning, information extraction, deep research, evidence collection, and report generation. The core Deep Researcher Agent employs four tools: reverse image search, metadata analysis, fact-checking databases, and verified news processing that extracts spatial, temporal, attribution, and motivational context. We demonstrate our approach on a challenge dataset sample involving complex multimedia content. Our system successfully verified content authenticity, extracted precise geolocation and timing information, and traced source attribution across multiple platforms, effectively addressing real-world multimedia verification scenarios.
comment: 33rd ACM International Conference on Multimedia (MM'25) Grand Challenge on Multimedia Verification
☆ BiFair: A Fairness-aware Training Framework for LLM-enhanced Recommender Systems via Bi-level Optimization
Large Language Model-enhanced Recommender Systems (LLM-enhanced RSs) have emerged as a powerful approach to improving recommendation quality by leveraging LLMs to generate item representations. Despite these advancements, the integration of LLMs raises severe fairness concerns. Existing studies reveal that LLM-based RSs exhibit greater unfairness than traditional RSs, yet fairness issues in LLM-enhanced RSs remain largely unexplored. In this paper, our empirical study reveals that while LLM-enhanced RSs improve fairness across item groups, a significant fairness gap persists. Further enhancement remains challenging due to the architectural differences and varying sources of unfairness inherent in LLM-enhanced RSs. To bridge this gap, we first decompose unfairness into i) \textit{prior unfairness} in LLM-generated representations and ii) \textit{training unfairness} in recommendation models. Then, we propose BiFair, a bi-level optimization-based fairness-aware training framework designed to mitigate both prior and training unfairness simultaneously. BiFair optimizes two sets of learnable parameters: LLM-generated representations and a trainable projector in the recommendation model, using a two-level nested optimization process. Additionally, we introduce an adaptive inter-group balancing mechanism, leveraging multi-objective optimization principles to dynamically balance fairness across item groups. Extensive experiments on three real-world datasets demonstrate that BiFair significantly mitigates unfairness and outperforms previous state-of-the-art methods.
☆ High-Resolution Sustain Pedal Depth Estimation from Piano Audio Across Room Acoustics
Piano sustain pedal detection has previously been approached as a binary on/off classification task, limiting its application in real-world piano performance scenarios where pedal depth significantly influences musical expression. This paper presents a novel approach for high-resolution estimation that predicts continuous pedal depth values. We introduce a Transformer-based architecture that not only matches state-of-the-art performance on the traditional binary classification task but also achieves high accuracy in continuous pedal depth estimation. Furthermore, by estimating continuous values, our model provides musically meaningful predictions for sustain pedal usage, whereas baseline models struggle to capture such nuanced expressions with their binary detection approach. Additionally, this paper investigates the influence of room acoustics on sustain pedal estimation using a synthetic dataset that includes varied acoustic conditions. We train our model with different combinations of room settings and test it in an unseen new environment using a "leave-one-out" approach. Our findings show that the two baseline models and ours are not robust to unseen room conditions. Statistical analysis further confirms that reverberation influences model predictions and introduces an overestimation bias.
♻ ☆ Long Context Modeling with Ranked Memory-Augmented Retrieval
Effective long-term memory management is crucial for language models handling extended contexts. We introduce a novel framework that dynamically ranks memory entries based on relevance. Unlike previous works, our model introduces a novel relevance scoring and a pointwise re-ranking model for key-value embeddings, inspired by learning-to-rank techniques in information retrieval. Enhanced Ranked Memory Augmented Retrieval ERMAR achieves state-of-the-art results on standard benchmarks.
♻ ☆ WARP: An Efficient Engine for Multi-Vector Retrieval SIGIR 2025
Multi-vector retrieval methods such as ColBERT and its recent variant, the ConteXtualized Token Retriever (XTR), offer high accuracy but face efficiency challenges at scale. To address this, we present WARP, a retrieval engine that substantially improves the efficiency of retrievers trained with the XTR objective through three key innovations: (1) WARP$_\text{SELECT}$ for dynamic similarity imputation; (2) implicit decompression, avoiding costly vector reconstruction during retrieval; and (3) a two-stage reduction process for efficient score aggregation. Combined with highly-optimized C++ kernels, our system reduces end-to-end latency compared to XTR's reference implementation by 41x, and achieves a 3x speedup over the ColBERTv2/PLAID engine, while preserving retrieval quality.
comment: Accepted at SIGIR 2025
♻ ☆ Towards Two-Stage Counterfactual Learning to Rank ICTIR 2025
Counterfactual learning to rank (CLTR) aims to learn a ranking policy from user interactions while correcting for the inherent biases in interaction data, such as position bias. Existing CLTR methods assume a single ranking policy that selects top-K ranking from the entire document candidate set. In real-world applications, the candidate document set is on the order of millions, making a single-stage ranking policy impractical. In order to scale to millions of documents, real-world ranking systems are designed in a two-stage fashion, with a candidate generator followed by a ranker. The existing CLTR method for a two-stage offline ranking system only considers the top-1 ranking set-up and only focuses on training the candidate generator, with the ranker fixed. A CLTR method for training both the ranker and candidate generator jointly is missing from the existing literature. In this paper, we propose a two-stage CLTR estimator that considers the interaction between the two stages and estimates the joint value of the two policies offline. In addition, we propose a novel joint optimization method to train the candidate and ranker policies, respectively. To the best of our knowledge, we are the first to propose a CLTR estimator and learning method for two-stage ranking. Experimental results on a semi-synthetic benchmark demonstrate the effectiveness of the proposed joint CLTR method over baselines.
comment: Accepted at ICTIR 2025 (co-located with SIGIR 2025)
Multimedia 3
☆ Multi-Modal Semantic Parsing for the Interpretation of Tombstone Inscriptions
Tombstones are historically and culturally rich artifacts, encapsulating individual lives, community memory, historical narratives and artistic expression. Yet, many tombstones today face significant preservation challenges, including physical erosion, vandalism, environmental degradation, and political shifts. In this paper, we introduce a novel multi-modal framework for tombstones digitization, aiming to improve the interpretation, organization and retrieval of tombstone content. Our approach leverages vision-language models (VLMs) to translate tombstone images into structured Tombstone Meaning Representations (TMRs), capturing both image and text information. To further enrich semantic parsing, we incorporate retrieval-augmented generation (RAG) for integrate externally dependent elements such as toponyms, occupation codes, and ontological concepts. Compared to traditional OCR-based pipelines, our method improves parsing accuracy from an F1 score of 36.1 to 89.5. We additionally evaluate the model's robustness across diverse linguistic and cultural inscriptions, and simulate physical degradation through image fusion to assess performance under noisy or damaged conditions. Our work represents the first attempt to formalize tombstone understanding using large vision-language models, presenting implications for heritage preservation.
comment: Accepted by ACMMM 2025
♻ ☆ AVTENet: A Human-Cognition-Inspired Audio-Visual Transformer-Based Ensemble Network for Video Deepfake Detection
The recent proliferation of hyper-realistic deepfake videos has drawn attention to the threat of audio and visual forgeries. Most previous studies on detecting artificial intelligence-generated fake videos only utilize visual modality or audio modality. While some methods exploit audio and visual modalities to detect forged videos, they have not been comprehensively evaluated on multimodal datasets of deepfake videos involving acoustic and visual manipulations, and are mostly based on convolutional neural networks with low detection accuracy. Considering that human cognition instinctively integrates multisensory information including audio and visual cues to perceive and interpret content and the success of transformer in various fields, this study introduces the audio-visual transformer-based ensemble network (AVTENet). This innovative framework tackles the complexities of deepfake technology by integrating both acoustic and visual manipulations to enhance the accuracy of video forgery detection. Specifically, the proposed model integrates several purely transformer-based variants that capture video, audio, and audio-visual salient cues to reach a consensus in prediction. For evaluation, we use the recently released benchmark multimodal audio-video FakeAVCeleb dataset. For a detailed analysis, we evaluate AVTENet, its variants, and several existing methods on multiple test sets of the FakeAVCeleb dataset. Experimental results show that the proposed model outperforms all existing methods and achieves state-of-the-art performance on Testset-I and Testset-II of the FakeAVCeleb dataset. We also compare AVTENet against humans in detecting video forgery. The results show that AVTENet significantly outperforms humans.
♻ ☆ Music102: An $D_{12}$-equivariant transformer for chord progression accompaniment
We present Music102, an advanced model aimed at enhancing chord progression accompaniment through a $D_{12}$-equivariant transformer. Inspired by group theory and symbolic music structures, Music102 leverages musical symmetry--such as transposition and reflection operations--integrating these properties into the transformer architecture. By encoding prior music knowledge, the model maintains equivariance across both melody and chord sequences. The POP909 dataset was employed to train and evaluate Music102, revealing significant improvements over the non-equivariant Music101 prototype Music101 in both weighted loss and exact accuracy metrics, despite using fewer parameters. This work showcases the adaptability of self-attention mechanisms and layer normalization to the discrete musical domain, addressing challenges in computational music analysis. With its stable and flexible neural framework, Music102 sets the stage for further exploration in equivariant music generation and computational composition tools, bridging mathematical theory with practical music performance.
comment: 10 pages, 3 figures
Information Retrieval 14
☆ Navigating Speech Recording Collections with AI-Generated Illustrations
Although the amount of available spoken content is steadily increasing, extracting information and knowledge from speech recordings remains challenging. Beyond enhancing traditional information retrieval methods such as speech search and keyword spotting, novel approaches for navigating and searching spoken content need to be explored and developed. In this paper, we propose a novel navigational method for speech archives that leverages recent advances in language and multimodal generative models. We demonstrate our approach with a Web application that organizes data into a structured format using interactive mind maps and image generation tools. The system is implemented using the TED-LIUM~3 dataset, which comprises over 2,000 speech transcripts and audio files of TED Talks. Initial user tests using a System Usability Scale (SUS) questionnaire indicate the application's potential to simplify the exploration of large speech collections.
☆ Cloud Digital Forensic Readiness: An Open Source Approach to Law Enforcement Request Management
Cloud Forensics presents a multi-jurisdictional challenge that may undermines the success of digital forensic investigations (DFIs). The growing volumes of domiciled and foreign law enforcement (LE) requests, the latency and complexity of formal channels for crossborder data access are challenging issues. In this paper, we first discuss major Cloud Service Providers (CSPs) transparency reports and law enforcement guidelines, then propose an abstract architecture for a Cloud Law Enforcement Requests Management System (CLERMS). A proof of concept of the proposed solution is developed, deployed and validated by two realistic scenarios, in addition to an economic estimation of its associated costs. Based on available open source components, our solution is for the benefit of both CSPs and Cloud Service Consumers (CSCs), and aims to enhance the due Cloud Digital Forensic Readiness (CDFR).
☆ Enhancing Learning Path Recommendation via Multi-task Learning
Personalized learning is a student-centered educational approach that adapts content, pace, and assessment to meet each learner's unique needs. As the key technique to implement the personalized learning, learning path recommendation sequentially recommends personalized learning items such as lectures and exercises. Advances in deep learning, particularly deep reinforcement learning, have made modeling such recommendations more practical and effective. This paper proposes a multi-task LSTM model that enhances learning path recommendation by leveraging shared information across tasks. The approach reframes learning path recommendation as a sequence-to-sequence (Seq2Seq) prediction problem, generating personalized learning paths from a learner's historical interactions. The model uses a shared LSTM layer to capture common features for both learning path recommendation and deep knowledge tracing, along with task-specific LSTM layers for each objective. To avoid redundant recommendations, a non-repeat loss penalizes repeated items within the recommended learning path. Experiments on the ASSIST09 dataset show that the proposed model significantly outperforms baseline methods for the learning path recommendation.
☆ CTR-Guided Generative Query Suggestion in Conversational Search
Generating effective query suggestions in conversational search requires aligning model outputs with user preferences, which is challenging due to sparse and noisy click signals. We propose GQS, a generative framework that integrates click modeling and preference optimization to enhance real-world user engagement. GQS consists of three key components: (1) a Multi-Source CTR Modeling module that captures diverse contextual signals to estimate fine-grained click-through rates; (2) a Diversity-Aware Preference Alignment strategy using CTR-weighted Direct Preference Optimization (DPO), which balances relevance and semantic diversity; and (3) a CTR-Calibrated Iterative Optimization process that jointly refines the CTR and generation models across training rounds. Experiments on two real-world tasks demonstrate that GQS outperforms strong baselines in CTR, relevance, and diversity.
☆ Leveraging Multimodal Data and Side Users for Diffusion Cross-Domain Recommendation
Cross-domain recommendation (CDR) aims to address the persistent cold-start problem in Recommender Systems. Current CDR research concentrates on transferring cold-start users' information from the auxiliary domain to the target domain. However, these systems face two main issues: the underutilization of multimodal data, which hinders effective cross-domain alignment, and the neglect of side users who interact solely within the target domain, leading to inadequate learning of the target domain's vector space distribution. To address these issues, we propose a model leveraging Multimodal data and Side users for diffusion Cross-domain recommendation (MuSiC). We first employ a multimodal large language model to extract item multimodal features and leverage a large language model to uncover user features using prompt learning without fine-tuning. Secondly, we propose the cross-domain diffusion module to learn the generation of feature vectors in the target domain. This approach involves learning feature distribution from side users and understanding the patterns in cross-domain transformation through overlapping users. Subsequently, the trained diffusion module is used to generate feature vectors for cold-start users in the target domain, enabling the completion of cross-domain recommendation tasks. Finally, our experimental evaluation of the Amazon dataset confirms that MuSiC achieves state-of-the-art performance, significantly outperforming all selected baselines. Our code is available: https://anonymous.4open.science/r/MuSiC-310A/.
☆ A Survey on Proactive Defense Strategies Against Misinformation in Large Language Models ACL 2025
The widespread deployment of large language models (LLMs) across critical domains has amplified the societal risks posed by algorithmically generated misinformation. Unlike traditional false content, LLM-generated misinformation can be self-reinforcing, highly plausible, and capable of rapid propagation across multiple languages, which traditional detection methods fail to mitigate effectively. This paper introduces a proactive defense paradigm, shifting from passive post hoc detection to anticipatory mitigation strategies. We propose a Three Pillars framework: (1) Knowledge Credibility, fortifying the integrity of training and deployed data; (2) Inference Reliability, embedding self-corrective mechanisms during reasoning; and (3) Input Robustness, enhancing the resilience of model interfaces against adversarial attacks. Through a comprehensive survey of existing techniques and a comparative meta-analysis, we demonstrate that proactive defense strategies offer up to 63\% improvement over conventional methods in misinformation prevention, despite non-trivial computational overhead and generalization challenges. We argue that future research should focus on co-designing robust knowledge foundations, reasoning certification, and attack-resistant interfaces to ensure LLMs can effectively counter misinformation across varied domains.
comment: Accepted by ACL 2025 Findings
☆ A Comparative Study of Specialized LLMs as Dense Retrievers
While large language models (LLMs) are increasingly deployed as dense retrievers, the impact of their domain-specific specialization on retrieval effectiveness remains underexplored. This investigation systematically examines how task-specific adaptations in LLMs influence their retrieval capabilities, an essential step toward developing unified retrievers capable of handling text, code, images, and multimodal content. We conduct extensive experiments with eight Qwen2.5 7B LLMs, including base, instruction-tuned, code/math-specialized, long reasoning, and vision-language models across zero-shot retrieval settings and the supervised setting. For the zero-shot retrieval settings, we consider text retrieval from the BEIR benchmark and code retrieval from the CoIR benchmark. Further, to evaluate supervised performance, all LLMs are fine-tuned on the MS MARCO dataset. We find that mathematical specialization and the long reasoning capability cause consistent degradation in three settings, indicating conflicts between mathematical reasoning and semantic matching. The vision-language model and code-specialized LLMs demonstrate superior zero-shot performance compared to other LLMs, even surpassing BM25 on the code retrieval task, and maintain comparable performance to base LLMs in supervised settings. These findings suggest promising directions for the unified retrieval task leveraging cross-domain and cross-modal fusion.
comment: Accepted by CCIR25 and published by Springer LNCS or LNAI
☆ Function-based Labels for Complementary Recommendation: Definition, Annotation, and LLM-as-a-Judge
Complementary recommendations enhance the user experience by suggesting items that are frequently purchased together while serving different functions from the query item. Inferring or evaluating whether two items have a complementary relationship requires complementary relationship labels; however, defining these labels is challenging because of the inherent ambiguity of such relationships. Complementary labels based on user historical behavior logs attempt to capture these relationships, but often produce inconsistent and unreliable results. Recent efforts have introduced large language models (LLMs) to infer these relationships. However, these approaches provide a binary classification without a nuanced understanding of complementary relationships. In this study, we address these challenges by introducing Function-Based Labels (FBLs), a novel definition of complementary relationships independent of user purchase logs and the opaque decision processes of LLMs. We constructed a human-annotated FBLs dataset comprising 2,759 item pairs and demonstrated that it covered possible item relationships and minimized ambiguity. We then evaluated whether some machine learning (ML) methods using annotated FBLs could accurately infer labels for unseen item pairs, and whether LLM-generated complementary labels align with human perception. Our results demonstrate that even with limited data, ML models, such as logistic regression and SVM achieve high macro-F1 scores (approximately 0.82). Furthermore, LLMs, such as gpt-4o-mini, demonstrated high consistency (0.989) and classification accuracy (0.849) under the detailed definition of FBLs, indicating their potential as effective annotators that mimic human judgment. Overall, our study presents FBLs as a clear definition of complementary relationships, enabling more accurate inferences and automated labeling of complementary recommendations.
☆ TayFCS: Towards Light Feature Combination Selection for Deep Recommender Systems
Feature interaction modeling is crucial for deep recommendation models. A common and effective approach is to construct explicit feature combinations to enhance model performance. However, in practice, only a small fraction of these combinations are truly informative. Thus it is essential to select useful feature combinations to reduce noise and manage memory consumption. While feature selection methods have been extensively studied, they are typically limited to selecting individual features. Extending these methods for high-order feature combination selection presents a significant challenge due to the exponential growth in time complexity when evaluating feature combinations one by one. In this paper, we propose $\textbf{TayFCS}$, a lightweight feature combination selection method that significantly improves model performance. Specifically, we propose the Taylor Expansion Scorer (TayScorer) module for field-wise Taylor expansion on the base model. Instead of evaluating all potential feature combinations' importance by repeatedly running experiments with feature adding and removal, this scorer only needs to approximate the importance based on their sub-components' gradients. This can be simply computed with one backward pass based on a trained recommendation model. To further reduce information redundancy among feature combinations and their sub-components, we introduce Logistic Regression Elimination (LRE), which estimates the corresponding information gain based on the model prediction performance. Experimental results on three benchmark datasets validate both the effectiveness and efficiency of our approach. Furthermore, online A/B test results demonstrate its practical applicability and commercial value.
☆ Continual Recommender Systems
Modern recommender systems operate in uniquely dynamic settings: user interests, item pools, and popularity trends shift continuously, and models must adapt in real time without forgetting past preferences. While existing tutorials on continual or lifelong learning cover broad machine learning domains (e.g., vision and graphs), they do not address recommendation-specific demands-such as balancing stability and plasticity per user, handling cold-start items, and optimizing recommendation metrics under streaming feedback. This tutorial aims to make a timely contribution by filling that gap. We begin by reviewing the background and problem settings, followed by a comprehensive overview of existing approaches. We then highlight recent efforts to apply continual learning to practical deployment environments, such as resource-constrained systems and sequential interaction settings. Finally, we discuss open challenges and future research directions. We expect this tutorial to benefit researchers and practitioners in recommender systems, data mining, AI, and information retrieval across academia and industry.
♻ ☆ Frustratingly Simple Retrieval Improves Challenging, Reasoning-Intensive Benchmarks
Retrieval-augmented Generation (RAG) has primarily been studied in limited settings, such as factoid question answering; more challenging, reasoning-intensive benchmarks have seen limited success from minimal RAG. In this work, we challenge this prevailing view on established, reasoning-intensive benchmarks: MMLU, MMLU Pro, AGI Eval, GPQA, and MATH. We identify a key missing component in prior work: a usable, web-scale datastore aligned with the breadth of pretraining data. To this end, we introduce CompactDS: a diverse, high-quality, web-scale datastore that achieves high retrieval accuracy and subsecond latency on a single-node. The key insights are (1) most web content can be filtered out without sacrificing coverage, and a compact, high-quality subset is sufficient; and (2) combining in-memory approximate nearest neighbor (ANN) retrieval and on-disk exact search balances speed and recall. Using CompactDS, we show that a minimal RAG pipeline achieves consistent accuracy improvements across all benchmarks and model sizes (8B--70B), with relative gains of 10% on MMLU, 33% on MMLU Pro, 14% on GPQA, and 19% on MATH. No single data source suffices alone, highlighting the importance of diversity of sources (web crawls, curated math, academic papers, textbooks). Finally, we show that our carefully designed in-house datastore matches or outperforms web search engines such as Google Search, as well as recently proposed, complex agent-based RAG systems--all while maintaining simplicity, reproducibility, and self-containment. We release CompactDS and our retrieval pipeline, supporting future research exploring retrieval-based AI systems.
comment: 33 pages, 2 figures, 27 tables
♻ ☆ FinBERT2: A Specialized Bidirectional Encoder for Bridging the Gap in Finance-Specific Deployment of Large Language Models
In natural language processing (NLP), the focus has shifted from encoder-only tiny language models like BERT to decoder-only large language models(LLMs) such as GPT-3. However, LLMs' practical application in the financial sector has revealed three limitations: (1) LLMs often perform worse than fine-tuned BERT on discriminative tasks despite costing much higher computational resources, such as market sentiment analysis in financial reports; (2) Application on generative tasks heavily relies on retrieval augmented generation (RAG) methods to provide current and specialized information, with general retrievers showing suboptimal performance on domain-specific retrieval tasks; (3) There are additional inadequacies in other feature-based scenarios, such as topic modeling. We introduce FinBERT2, a specialized bidirectional encoder pretrained on a high-quality, financial-specific corpus of 32b tokens. This represents the largest known Chinese financial pretraining corpus for models of this parameter size. As a better backbone, FinBERT2 can bridge the gap in the financial-specific deployment of LLMs through the following achievements: (1) Discriminative fine-tuned models (Fin-Labelers) outperform other (Fin)BERT variants by 0.4%-3.3% and leading LLMs by 9.7%-12.3% on average across five financial classification tasks. (2) Contrastive fine-tuned models (Fin-Retrievers) outperform both open-source (e.g., +6.8\% avg improvement over BGE-base-zh) and proprietary (e.g., +4.2\% avg improvement over OpenAI's text-embedding-3-large) embedders across five financial retrieval tasks; (3) Building on FinBERT2 variants, we construct the Fin-TopicModel, which enables superior clustering and topic representation for financial titles. Our work revisits financial BERT models through comparative analysis with contemporary LLMs and offers practical insights for effectively utilizing FinBERT in the LLMs era.
♻ ☆ From Prompting to Alignment: A Generative Framework for Query Recommendation
In modern search systems, search engines often suggest relevant queries to users through various panels or components, helping refine their information needs. Traditionally, these recommendations heavily rely on historical search logs to build models, which suffer from cold-start or long-tail issues. Furthermore, tasks such as query suggestion, completion or clarification are studied separately by specific design, which lacks generalizability and hinders adaptation to novel applications. Despite recent attempts to explore the use of LLMs for query recommendation, these methods mainly rely on the inherent knowledge of LLMs or external sources like few-shot examples, retrieved documents, or knowledge bases, neglecting the importance of the calibration and alignment with user feedback, thus limiting their practical utility. To address these challenges, we first propose a general Generative Query Recommendation (GQR) framework that aligns LLM-based query generation with user preference. Specifically, we unify diverse query recommendation tasks by a universal prompt framework, leveraging the instruct-following capability of LLMs for effective generation. Secondly, we align LLMs with user feedback via presenting a CTR-alignment framework, which involves training a query-wise CTR predictor as a process reward model and employing list-wise preference alignment to maximize the click probability of the generated query list. Furthermore, recognizing the inconsistency between LLM knowledge and proactive search intents arising from the separation of user-initiated queries from models, we align LLMs with user initiative via retrieving co-occurrence queries as side information when historical logs are available.
♻ ☆ Explainable Coarse-to-Fine Ancient Manuscript Duplicates Discovery
Ancient manuscripts are the primary source of ancient linguistic corpora. However, many ancient manuscripts exhibit duplications due to unintentional repeated publication or deliberate forgery. The Dead Sea Scrolls, for example, include counterfeit fragments, whereas Oracle Bones (OB) contain both republished materials and fabricated specimens. Identifying ancient manuscript duplicates is of great significance for both archaeological curation and ancient history study. In this work, we design a progressive OB duplicate discovery framework that combines unsupervised low-level keypoints matching with high-level text-centric content-based matching to refine and rank the candidate OB duplicates with semantic awareness and interpretability. We compare our model with state-of-the-art content-based image retrieval and image matching methods, showing that our model yields comparable recall performance and the highest simplified mean reciprocal rank scores for both Top-5 and Top-15 retrieval results, and with significantly accelerated computation efficiency. We have discovered over 60 pairs of new OB duplicates in real-world deployment, which were missed by domain experts for decades. Code, model and real-world results are available at: https://github.com/cszhangLMU/OBD-Finder/.
comment: Explainable Coarse-to-Fine Ancient Manuscript Duplicates Discovery, with Oracle Bones as a Case Study
Multimedia 4
☆ Consistent and Invariant Generalization Learning for Short-video Misinformation Detection ACM MM 2025
Short-video misinformation detection has attracted wide attention in the multi-modal domain, aiming to accurately identify the misinformation in the video format accompanied by the corresponding audio. Despite significant advancements, current models in this field, trained on particular domains (source domains), often exhibit unsatisfactory performance on unseen domains (target domains) due to domain gaps. To effectively realize such domain generalization on the short-video misinformation detection task, we propose deep insights into the characteristics of different domains: (1) The detection on various domains may mainly rely on different modalities (i.e., mainly focusing on videos or audios). To enhance domain generalization, it is crucial to achieve optimal model performance on all modalities simultaneously. (2) For some domains focusing on cross-modal joint fraud, a comprehensive analysis relying on cross-modal fusion is necessary. However, domain biases located in each modality (especially in each frame of videos) will be accumulated in this fusion process, which may seriously damage the final identification of misinformation. To address these issues, we propose a new DOmain generalization model via ConsisTency and invariance learning for shORt-video misinformation detection (named DOCTOR), which contains two characteristic modules: (1) We involve the cross-modal feature interpolation to map multiple modalities into a shared space and the interpolation distillation to synchronize multi-modal learning; (2) We design the diffusion model to add noise to retain core features of multi modal and enhance domain invariant features through cross-modal guided denoising. Extensive experiments demonstrate the effectiveness of our proposed DOCTOR model. Our code is public available at https://github.com/ghh1125/DOCTOR.
comment: Accepted to ACM MM 2025,15 pages, 16figures
☆ The shortcomings of video conferencing technology, methods for revealing them, and emerging XR solutions
Video conferencing has become a central part of our daily lives, thanks to the COVID-19 pandemic. Unfortunately, so have its many limitations, resulting in poor support for communicative and social behavior and ultimately, Zoom fatigue. New technologies will be required to address these limitations, including many drawn from mixed reality (XR). In this paper, our goals are to equip and encourage future researchers to develop and test such technologies. Toward this end, we first survey research on the shortcomings of video conferencing systems, as defined before and after the pandemic. We then consider the methods that research uses to evaluate support for communicative behavior, and argue that those same methods should be employed in identifying, improving, and validating promising video conferencing technologies. Next, we survey emerging XR solutions to video conferencing's limitations, most off which do not employ head-mounted displays.
☆ From Query to Explanation: Uni-RAG for Multi-Modal Retrieval-Augmented Learning in STEM
In AI-facilitated teaching, leveraging various query styles to interpret abstract educational content is crucial for delivering effective and accessible learning experiences. However, existing retrieval systems predominantly focus on natural text-image matching and lack the capacity to address the diversity and ambiguity inherent in real-world educational scenarios. To address this limitation, we develop a lightweight and efficient multi-modal retrieval module, named Uni-Retrieval, which extracts query-style prototypes and dynamically matches them with tokens from a continually updated Prompt Bank. This Prompt Bank encodes and stores domain-specific knowledge by leveraging a Mixture-of-Expert Low-Rank Adaptation (MoE-LoRA) module and can be adapted to enhance Uni-Retrieval's capability to accommodate unseen query types at test time. To enable natural language educational content generation, we integrate the original Uni-Retrieval with a compact instruction-tuned language model, forming a complete retrieval-augmented generation pipeline named Uni-RAG. Given a style-conditioned query, Uni-RAG first retrieves relevant educational materials and then generates human-readable explanations, feedback, or instructional content aligned with the learning objective. Experimental results on SER and other multi-modal benchmarks show that Uni-RAG outperforms baseline retrieval and RAG systems in both retrieval accuracy and generation quality, while maintaining low computational cost. Our framework provides a scalable, pedagogically grounded solution for intelligent educational systems, bridging retrieval and generation to support personalized, explainable, and efficient learning assistance across diverse STEM scenarios.
♻ ☆ AVE Speech: A Comprehensive Multi-Modal Dataset for Speech Recognition Integrating Audio, Visual, and Electromyographic Signals
The global aging population faces considerable challenges, particularly in communication, due to the prevalence of hearing and speech impairments. To address these, we introduce the AVE speech, a comprehensive multi-modal dataset for speech recognition tasks. The dataset includes a 100-sentence Mandarin corpus with audio signals, lip-region video recordings, and six-channel electromyography (EMG) data, collected from 100 participants. Each subject read the entire corpus ten times, with each sentence averaging approximately two seconds in duration, resulting in over 55 hours of multi-modal speech data per modality. Experiments demonstrate that combining these modalities significantly improves recognition performance, particularly in cross-subject and high-noise environments. To our knowledge, this is the first publicly available sentence-level dataset integrating these three modalities for large-scale Mandarin speech recognition. We expect this dataset to drive advancements in both acoustic and non-acoustic speech recognition research, enhancing cross-modal learning and human-machine interaction.
comment: The paper has been accepted by IEEE Transactions on Human-Machine Systems
Information Retrieval 17
☆ Beyond classical and contemporary models: a transformative ai framework for student dropout prediction in distance learning using rag, prompt engineering, and cross-modal fusion
Student dropout in distance learning remains a critical challenge, with profound societal and economic consequences. While classical machine learning models leverage structured socio-demographic and behavioral data, they often fail to capture the nuanced emotional and contextual factors embedded in unstructured student interactions. This paper introduces a transformative AI framework that redefines dropout prediction through three synergistic innovations: Retrieval-Augmented Generation (RAG) for domain-specific sentiment analysis, prompt engineering to decode academic stressors, and cross-modal attention fusion to dynamically align textual, behavioral, and socio-demographic insights. By grounding sentiment analysis in a curated knowledge base of pedagogical content, our RAG-enhanced BERT model interprets student comments with unprecedented contextual relevance, while optimized prompts isolate indicators of academic distress (e.g., "isolation," "workload anxiety"). A cross-modal attention layer then fuses these insights with temporal engagement patterns, creating holistic risk profiles. Evaluated on a longitudinal dataset of 4 423 students, the framework achieves 89% accuracy and an F1-score of 0.88, outperforming conventional models by 7% and reducing false negatives by 21%. Beyond prediction, the system generates interpretable interventions by retrieving contextually aligned strategies (e.g., mentorship programs for isolated learners). This work bridges the gap between predictive analytics and actionable pedagogy, offering a scalable solution to mitigate dropout risks in global education systems
comment: 10 pages, 5 figures, 5 tables. Submitted to the 38th Canadian Conference on Artificial Intelligence (Canadian AI 2025)
☆ Efficient and Effective Query Context-Aware Learning-to-Rank Model for Sequential Recommendation
Modern sequential recommender systems commonly use transformer-based models for next-item prediction. While these models demonstrate a strong balance between efficiency and quality, integrating interleaving features - such as the query context (e.g., browse category) under which next-item interactions occur - poses challenges. Effectively capturing query context is crucial for refining ranking relevance and enhancing user engagement, as it provides valuable signals about user intent within a session. Unlike an item's features, query context is not temporally aligned with the item sequence, making its incorporation into transformers challenging and error-prone. This paper analyzes different strategies for incorporating query context into transformers trained with a causal language modeling procedure as a case study. We propose a new method that effectively fuses the item sequence with query context within the attention mechanism. Through extensive offline and online experiments on a large-scale online platform and open datasets, we present evidence that our proposed method is an effective approach for integrating query context to improve model ranking quality in terms of relevance and diversity.
☆ Ranking-based Fusion Algorithms for Extreme Multi-label Text Classification (XMTC)
In the context of Extreme Multi-label Text Classification (XMTC), where labels are assigned to text instances from a large label space, the long-tail distribution of labels presents a significant challenge. Labels can be broadly categorized into frequent, high-coverage \textbf{head labels} and infrequent, low-coverage \textbf{tail labels}, complicating the task of balancing effectiveness across all labels. To address this, combining predictions from multiple retrieval methods, such as sparse retrievers (e.g., BM25) and dense retrievers (e.g., fine-tuned BERT), offers a promising solution. The fusion of \textit{sparse} and \textit{dense} retrievers is motivated by the complementary ranking characteristics of these methods. Sparse retrievers compute relevance scores based on high-dimensional, bag-of-words representations, while dense retrievers utilize approximate nearest neighbor (ANN) algorithms on dense text and label embeddings within a shared embedding space. Rank-based fusion algorithms leverage these differences by combining the precise matching capabilities of sparse retrievers with the semantic richness of dense retrievers, thereby producing a final ranking that improves the effectiveness across both head and tail labels.
☆ Agent-Based Detection and Resolution of Incompleteness and Ambiguity in Interactions with Large Language Models
Many of us now treat LLMs as modern-day oracles asking it almost any kind of question. However, consulting an LLM does not have to be a single turn activity. But long multi-turn interactions can get tedious if it is simply to clarify contextual information that can be arrived at through reasoning. In this paper, we examine the use of agent-based architecture to bolster LLM-based Question-Answering systems with additional reasoning capabilities. We examine the automatic resolution of potential incompleteness or ambiguities in questions by transducers implemented using LLM-based agents. We focus on several benchmark datasets that are known to contain questions with these deficiencies to varying degrees. We equip different LLMs (GPT-3.5-Turbo and Llama-4-Scout) with agents that act as specialists in detecting and resolving deficiencies of incompleteness and ambiguity. The agents are implemented as zero-shot ReAct agents. Rather than producing an answer in a single step, the model now decides between 3 actions a) classify b) resolve c) answer. Action a) decides if the question is incomplete, ambiguous, or normal. Action b) determines if any deficiencies identified can be resolved. Action c) answers the resolved form of the question. We compare the use of LLMs with and without the use of agents with these components. Our results show benefits of agents with transducer 1) A shortening of the length of interactions with human 2) An improvement in the answer quality and 3) Explainable resolution of deficiencies in the question. On the negative side we find while it may result in additional LLM invocations and in some cases, increased latency. But on tested datasets, the benefits outweigh the costs except when questions already have sufficient context. Suggesting the agent-based approach could be a useful mechanism to harness the power of LLMs to develop more robust QA systems.
comment: 14 pages. arXiv admin note: text overlap with arXiv:2503.17936
☆ GENPLUGIN: A Plug-and-Play Framework for Long-Tail Generative Recommendation with Exposure Bias Mitigation
Generative recommendation (GenRec) offers LLM integration, reduced embedding costs, and eliminates per-candidate scoring, attracting great attention. Despite its promising performance, this study reveals that it suffers from generation exposure bias and poor long-tail item generalization, two critical limitations overlooked by prior works on GenRec. To address these, we propose GENPLUGIN, a plug-and-play framework featuring a dual-encoder, shared-decoder architecture. During pre-training, it aligns language and ID views via contrastive learning, harmonizing item representations across two complementary views. Besides, GENPLUGIN uses a novel training strategy that probabilistically substitutes ground-truth item ID tokens with predictions from the language-semantics encoder, alleviating exposure bias. To improve long-tail generative recommendation, we propose a retrieval-based data augmentation mechanism. It fine-tunes the decoder of GENPLUGIN to endow GENPLUGIN with the ability to use relevant users w.r.t. contexts or collaborative information to augment the generation of item ID tokens in long-tail recommendation scenarios. We have plugged GENPLUGIN into several representative GenRec models and the extensive experiments demonstrate that GENPLUGIN can notably mitigate generation exposure bias during item ID generation while significantly improving the quality of long-tail item recommendation.
comment: 16 pages, 8 figures
☆ A Multistakeholder Approach to Value-Driven Co-Design of Recommender System Evaluation Metrics in Digital Archives RecSys'25
This paper presents the first multistakeholder approach for translating diverse stakeholder values into an evaluation metric setup for Recommender Systems (RecSys) in digital archives. While commercial platforms mainly rely on engagement metrics, cultural heritage domains require frameworks that balance competing priorities among archivists, platform owners, researchers, and other stakeholders. To address this challenge, we conducted high-profile focus groups (5 groups x 5 persons) with upstream, provider, system, consumer, and downstream stakeholders, identifying value priorities across critical dimensions: visibility/representation, expertise adaptation, and transparency/trust. Our analysis shows that stakeholder concerns naturally align with four sequential research funnel stages: discovery, interaction, integration, and impact. The resulting framework addresses domain-specific challenges including collection representation imbalances, non-linear research patterns, and tensions between specialized expertise and broader accessibility. We propose tailored metrics for each stage in this research journey, such as research path quality for discovery, contextual appropriateness for interaction, metadata-weighted relevance for integration, and cross-stakeholder value alignment for impact assessment. Our contributions extend beyond digital archives to the broader RecSys community, offering transferable evaluation approaches for domains where value emerges through sustained engagement rather than immediate consumption.
comment: Accepted at RecSys'25
☆ Exploring the Effect of Context-Awareness and Popularity Calibration on Popularity Bias in POI Recommendations RecSys 2025
Point-of-interest (POI) recommender systems help users discover relevant locations, but their effectiveness is often compromised by popularity bias, which disadvantages less popular, yet potentially meaningful places. This paper addresses this challenge by evaluating the effectiveness of context-aware models and calibrated popularity techniques as strategies for mitigating popularity bias. Using four real-world POI datasets (Brightkite, Foursquare, Gowalla, and Yelp), we analyze the individual and combined effects of these approaches on recommendation accuracy and popularity bias. Our results reveal that context-aware models cannot be considered a uniform solution, as the models studied exhibit divergent impacts on accuracy and bias. In contrast, calibration techniques can effectively align recommendation popularity with user preferences, provided there is a careful balance between accuracy and bias mitigation. Notably, the combination of calibration and context-awareness yields recommendations that balance accuracy and close alignment with the users' popularity profiles, i.e., popularity calibration.
comment: Accepted at RecSys 2025
☆ Explainable Information Retrieval in the Audit Domain SIGIR 2025
Conversational agents such as Microsoft Copilot and Google Gemini assist users with complex search tasks but often generate misleading or fabricated references. This undermines trust, particularly in high-stakes domains such as medicine and finance. Explainable information retrieval (XIR) aims to address this by making search results more transparent and interpretable. While most XIR research is domain-agnostic, this paper focuses on auditing -- a critical yet underexplored area. We argue that XIR systems can support auditors in completing their complex task. We outline key challenges and future research directions to advance XIR in this domain.
comment: Extended abstract accepted at the Workshop on Explainability in Information Retrieval (WExIR), co-located with SIGIR 2025
☆ Exploring LLM Capabilities in Extracting DCAT-Compatible Metadata for Data Cataloging
Efficient data exploration is crucial as data becomes increasingly important for accelerating processes, improving forecasts and developing new business models. Data consumers often spend 25-98 % of their time searching for suitable data due to the exponential growth, heterogeneity and distribution of data. Data catalogs can support and accelerate data exploration by using metadata to answer user queries. However, as metadata creation and maintenance is often a manual process, it is time-consuming and requires expertise. This study investigates whether LLMs can automate metadata maintenance of text-based data and generate high-quality DCAT-compatible metadata. We tested zero-shot and few-shot prompting strategies with LLMs from different vendors for generating metadata such as titles and keywords, along with a fine-tuned model for classification. Our results show that LLMs can generate metadata comparable to human-created content, particularly on tasks that require advanced semantic understanding. Larger models outperformed smaller ones, and fine-tuning significantly improves classification accuracy, while few-shot prompting yields better results in most cases. Although LLMs offer a faster and reliable way to create metadata, a successful application requires careful consideration of task-specific criteria and domain context.
☆ Modeling Item-Level Dynamic Variability with Residual Diffusion for Bundle Recommendation
Existing solutions for bundle recommendation(BR) have achieved remarkable effectiveness for predicting the user's preference for prebuilt bundles. However, bundle-item(B-I) affiliation will vary dynamically in real scenarios. For example, a bundle themed as 'casual outfit', may add 'hat' or remove 'watch' due to factors such as seasonal variations, changes in user pes or inventory adjustments. Our empirical study demonstrates that the performance of mainstream BR models will fluctuate or even decline regarding item-level variability. This paper makes the first attempt to referencaddress the above problem and proposes a novel Residual Diffusion for Bundle Recommendation(RDiffBR) as a model-agnostic generative framework which can assist a BR model in adapting this scenario. During the initial training of the BR model, RDiffBR employs a residual diffusion model to process the item-level bundle embeddings which are generated by BR model to represent bundle theme via a forward-reverse process. In the inference stage, RDiffBR reverses item-level bundle embeddings obtained by the well-trained bundle model under B-I variability scenarios to generate the effective item-level bundle embeddings. In particular, the residual connection in our residual approximator significantly enhances item-level bundle embeddings generation ability of BR models. Experiments on six BR models and four public datasets from different domains show that RDiffBR improves the performance of Recall and NDCG of backbone BR models by up to 23%, while only increased training time about 4%.Codes and datasets are available at https://anonymous.4open.science/r/RDiffBR.
♻ ☆ Towards Fair RAG: On the Impact of Fair Ranking in Retrieval-Augmented Generation ICTIR 2025
Despite the central role of retrieval in retrieval-augmented generation (RAG) systems, much of the existing research on RAG overlooks the well-established field of fair ranking and fails to account for the interests of all stakeholders involved. In this paper, we conduct the first systematic evaluation of RAG systems that integrate fairness-aware rankings, addressing both ranking fairness and attribution fairness, which ensures equitable exposure of the sources cited in the generated content. Our evaluation focuses on measuring item-side fairness, specifically the fair exposure of relevant items retrieved by RAG systems, and investigates how this fairness impacts both the effectiveness of the systems and the attribution of sources in the generated output that users ultimately see. By experimenting with twelve RAG models across seven distinct tasks, we show that incorporating fairness-aware retrieval often maintains or even enhances both ranking quality and generation quality, countering the common belief that fairness compromises system performance. Additionally, we demonstrate that fair retrieval practices lead to more balanced attribution in the final responses, ensuring that the generator fairly cites the sources it relies on. Our findings underscore the importance of item-side fairness in retrieval and generation, laying the foundation for responsible and equitable RAG systems and guiding future research in fair ranking and attribution.
comment: ICTIR 2025 (Oral); NeurIPS 2024 AFME Workshop (Spotlight)
♻ ☆ Tip of the Tongue Query Elicitation for Simulated Evaluation SIGIR 2025
Tip-of-the-tongue (TOT) search occurs when a user struggles to recall a specific identifier, such as a document title. While common, existing search systems often fail to effectively support TOT scenarios. Research on TOT retrieval is further constrained by the challenge of collecting queries, as current approaches rely heavily on community question-answering (CQA) websites, leading to labor-intensive evaluation and domain bias. To overcome these limitations, we introduce two methods for eliciting TOT queries - leveraging large language models (LLMs) and human participants - to facilitate simulated evaluations of TOT retrieval systems. Our LLM-based TOT user simulator generates synthetic TOT queries at scale, achieving high correlations with how CQA-based TOT queries rank TOT retrieval systems when tested in the Movie domain. Additionally, these synthetic queries exhibit high linguistic similarity to CQA-derived queries. For human-elicited queries, we developed an interface that uses visual stimuli to place participants in a TOT state, enabling the collection of natural queries. In the Movie domain, system rank correlation and linguistic similarity analyses confirm that human-elicited queries are both effective and closely resemble CQA-based queries. These approaches reduce reliance on CQA-based data collection while expanding coverage to underrepresented domains, such as Landmark and Person. LLM-elicited queries for the Movie, Landmark, and Person domains have been released as test queries in the TREC 2024 TOT track, with human-elicited queries scheduled for inclusion in the TREC 2025 TOT track. Additionally, we provide source code for synthetic query generation and the human query collection interface, along with curated visual stimuli used for eliciting TOT queries.
comment: SIGIR 2025
♻ ☆ DTN: Deep Multiple Task-specific Feature Interactions Network for Multi-Task Recommendation
Neural-based multi-task learning (MTL) has been successfully applied to many recommendation applications. However, these MTL models (e.g., MMoE, PLE) did not consider feature interaction during the optimization, which is crucial for capturing complex high-order features and has been widely used in ranking models for real-world recommender systems. Moreover, through feature importance analysis across various tasks in MTL, we have observed an interesting divergence phenomenon that the same feature can have significantly different importance across different tasks in MTL. To address these issues, we propose Deep Multiple Task-specific Feature Interactions Network (DTN) with a novel model structure design. DTN introduces multiple diversified task-specific feature interaction methods and task-sensitive network in MTL networks, enabling the model to learn task-specific diversified feature interaction representations, which improves the efficiency of joint representation learning in a general setup. We applied DTN to our company's real-world E-commerce recommendation dataset, which consisted of over 6.3 billion samples, the results demonstrated that DTN significantly outperformed state-of-the-art MTL models. Moreover, during online evaluation of DTN in a large-scale E-commerce recommender system, we observed a 3.28% in clicks, a 3.10% increase in orders and a 2.70% increase in GMV (Gross Merchandise Value) compared to the state-of-the-art MTL models. Finally, extensive offline experiments conducted on public benchmark datasets demonstrate that DTN can be applied to various scenarios beyond recommendations, enhancing the performance of ranking models.
♻ ☆ Deep Retrieval at CheckThat! 2025: Identifying Scientific Papers from Implicit Social Media Mentions via Hybrid Retrieval and Re-Ranking
We present the methodology and results of the Deep Retrieval team for subtask 4b of the CLEF CheckThat! 2025 competition, which focuses on retrieving relevant scientific literature for given social media posts. To address this task, we propose a hybrid retrieval pipeline that combines lexical precision, semantic generalization, and deep contextual re-ranking, enabling robust retrieval that bridges the informal-to-formal language gap. Specifically, we combine BM25-based keyword matching with a FAISS vector store using a fine-tuned INF-Retriever-v1 model for dense semantic retrieval. BM25 returns the top 30 candidates, and semantic search yields 100 candidates, which are then merged and re-ranked via a large language model (LLM)-based cross-encoder. Our approach achieves a mean reciprocal rank at 5 (MRR@5) of 76.46% on the development set and 66.43% on the hidden test set, securing the 1st position on the development leaderboard and ranking 3rd on the test leaderboard (out of 31 teams), with a relative performance gap of only 2 percentage points compared to the top-ranked system. We achieve this strong performance by running open-source models locally and without external training data, highlighting the effectiveness of a carefully designed and fine-tuned retrieval pipeline.
♻ ☆ Recommender systems, stigmergy, and the tyranny of popularity
Scientific recommender systems, such as Google Scholar and Web of Science, are essential tools for discovery. Search algorithms that power work through stigmergy, a collective intelligence mechanism that surfaces useful paths through repeated engagement. While generally effective, this "rich-get-richer" dynamic results in a small number of high-profile papers that dominate visibility. This essay argues argue that these algorithm over-reliance on popularity fosters intellectual homogeneity and exacerbates structural inequities, stifling innovative and diverse perspectives critical for scientific progress. We propose an overhaul of search platforms to incorporate user-specific calibration, allowing researchers to manually adjust the weights of factors like popularity, recency, and relevance. We also advise platform developers on how text embeddings and LLMs could be implemented in ways that increase user autonomy. While our suggestions are particularly pertinent to aligning recommender systems with scientific values, these ideas are broadly applicable to information access systems in general. Designing platforms that increase user autonomy is an important step toward more robust and dynamic information
♻ ☆ EraRAG: Efficient and Incremental Retrieval Augmented Generation for Growing Corpora
Graph-based Retrieval-Augmented Generation (Graph-RAG) enhances large language models (LLMs) by structuring retrieval over an external corpus. However, existing approaches typically assume a static corpus, requiring expensive full-graph reconstruction whenever new documents arrive, limiting their scalability in dynamic, evolving environments. To address these limitations, we introduce EraRAG, a novel multi-layered Graph-RAG framework that supports efficient and scalable dynamic updates. Our method leverages hyperplane-based Locality-Sensitive Hashing (LSH) to partition and organize the original corpus into hierarchical graph structures, enabling efficient and localized insertions of new data without disrupting the existing topology. The design eliminates the need for retraining or costly recomputation while preserving high retrieval accuracy and low latency. Experiments on large-scale benchmarks demonstrate that EraRag achieves up to an order of magnitude reduction in update time and token consumption compared to existing Graph-RAG systems, while providing superior accuracy performance. This work offers a practical path forward for RAG systems that must operate over continually growing corpora, bridging the gap between retrieval efficiency and adaptability. Our code and data are available at https://github.com/EverM0re/EraRAG-Official.
comment: Under review
♻ ☆ Performance-Driven QUBO for Recommender Systems on Quantum Annealers
We propose Counterfactual Analysis Quadratic Unconstrained Binary Optimization (CAQUBO) to solve QUBO problems for feature selection in recommender systems. CAQUBO leverages counterfactual analysis to measure the impact of individual features and feature combinations on model performance and employs the measurements to construct the coefficient matrix for a quantum annealer to select the optimal feature combinations for recommender systems, thereby improving their final recommendation performance. By establishing explicit connections between features and the recommendation performance, the proposed approach demonstrates superior performance compared to the state-of-the-art quantum annealing methods. Extensive experiments indicate that integrating quantum computing with counterfactual analysis holds great promise for addressing these challenges.
Multimedia 4
☆ Assessing the Viability of Wave Field Synthesis in VR-Based Cognitive Research
This paper investigates the viability of Wave Field Synthesis (WFS) for enhancing auditory immersion in VR-based cognitive research. While Virtual Reality (VR) offers significant advantages for studying human perception and behavior, auditory cues are often underutilized. WFS, an advanced audio rendering technique, can create highly realistic and spatially accurate soundscapes, potentially increasing ecological validity. This study evaluates WFS by implementing a sample experiment where participants localize static and moving sound sources in both a WFS-rendered environment and a conventional stereo headphone setup. The research explores the impact of virtual environments, sound types, and durations on localization accuracy and search behavior. Findings indicate that while stereo setups can achieve higher accuracy, WFS provides a more natural and intuitive auditory experience, particularly for directional cues. The study also highlights limitations of current WFS systems, such as the lack of height localization, occlusion simulation, and user-dependent optimization, which affect performance, especially for centrally located sound sources. Despite these challenges, WFS shows promise for specialized auditory perception research, particularly for complex soundscapes where directional information is paramount.
comment: 35 pages
☆ Unlearning the Noisy Correspondence Makes CLIP More Robust ICCV 2025
The data appetite for Vision-Language Models (VLMs) has continuously scaled up from the early millions to billions today, which faces an untenable trade-off with data quality and inevitably introduces Noisy Correspondence (NC) samples. Undoubtedly, such semantically unrelated data significantly impairs the performance of VLMs. Previous efforts mainly address this challenge by estimating refined alignment for more precise guidance. However, such resource-intensive pipelines that train VLMs from scratch struggle to meet realistic data demands. In this paper, we present a brand new perspective that seeks to directly eliminate the harmful effects of NC in pre-trained VLMs. Specifically, we propose NCU, a Noisy Correspondence Unlearning fine-tuning framework that efficiently enhances VLMs' robustness by forgetting learned noisy knowledge. The key to NCU is learning the hardest negative information, which can provide explicit unlearning direction for both false positives and false negatives. Such twin goals unlearning process can be formalized into one unified optimal transport objective for fast fine-tuning. We validate our approach with the prevailing CLIP model over various downstream tasks. Remarkably, NCU surpasses the robust pre-trained method on zero-shot transfer while with lower computational overhead. The code will be released upon acceptance.
comment: ICCV 2025
☆ Gaze and Glow: Exploring Editing Processes on Social Media through Interactive Exhibition
We present Gaze and Glow, an interactive installation that reveals the often-invisible efforts of social media editing. Through narrative personas, experimental videos, and sensor-based interactions, the installation explores how audience attention shapes users' editing practices and emotional experiences. Deployed in a two-month public exhibition, Gaze and Glow engaged viewers and elicited responses. Reflexive thematic analysis of audience feedback highlights how making editing visible prompts new reflections on authenticity, agency, and performativity. We discuss implications for designing interactive systems that support selective memory, user-controlled visibility, and critical engagement with everyday digital self-presentation.
comment: 6 pages, 6 figures, to be published in DIS 2025 (Provocations and Works in Progress)
♻ ☆ Multichannel Steganography: A Provably Secure Hybrid Steganographic Model for Secure Communication
Secure covert communication in hostile environments requires simultaneously achieving invisibility, provable security guarantees, and robustness against informed adversaries. This paper presents a novel hybrid steganographic framework that unites cover synthesis and cover modification within a unified multichannel protocol. A secret-seeded PRNG drives a lightweight Markov-chain generator to produce contextually plausible cover parameters, which are then masked with the payload and dispersed across independent channels. The masked bit-vector is imperceptibly embedded into conventional media via a variance-aware least-significant-bit algorithm, ensuring that statistical properties remain within natural bounds. We formalize a multichannel adversary model (MC-ATTACK) and prove that, under standard security assumptions, the adversary's distinguishing advantage is negligible, thereby guaranteeing both confidentiality and integrity. Empirical results corroborate these claims: local-variance-guided embedding yields near-lossless extraction (mean BER $<5\times10^{-3}$, correlation $>0.99$) with minimal perceptual distortion (PSNR $\approx100$,dB, SSIM $>0.99$), while key-based masking drives extraction success to zero (BER $\approx0.5$) for a fully informed adversary. Comparative analysis demonstrates that purely distortion-free or invertible schemes fail under the same threat model, underscoring the necessity of hybrid designs. The proposed approach advances high-assurance steganography by delivering an efficient, provably secure covert channel suitable for deployment in high-surveillance networks.
comment: 22 pages, 15 figures, 4 algorithms. This version is a preprint uploaded to arXiv